WO2012109384A1 - Suppression de bruit combinée et signaux hors emplacement - Google Patents

Suppression de bruit combinée et signaux hors emplacement Download PDF

Info

Publication number
WO2012109384A1
WO2012109384A1 PCT/US2012/024370 US2012024370W WO2012109384A1 WO 2012109384 A1 WO2012109384 A1 WO 2012109384A1 US 2012024370 W US2012024370 W US 2012024370W WO 2012109384 A1 WO2012109384 A1 WO 2012109384A1
Authority
WO
WIPO (PCT)
Prior art keywords
gain
noise
suppression
banded
recited
Prior art date
Application number
PCT/US2012/024370
Other languages
English (en)
Inventor
Glenn N. Dickins
Timothy J. NEAL
Mark S. Vinton
Original Assignee
Dolby Laboratories Licensing Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Dolby Laboratories Licensing Corporation filed Critical Dolby Laboratories Licensing Corporation
Priority to JP2013553528A priority Critical patent/JP6002690B2/ja
Priority to CN201280008266.XA priority patent/CN103348408B/zh
Priority to EP12707412.8A priority patent/EP2673777B1/fr
Publication of WO2012109384A1 publication Critical patent/WO2012109384A1/fr
Priority to US13/964,037 priority patent/US9173025B2/en

Links

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G10L21/0216—Noise filtering characterised by the method used for estimating noise
    • G10L2021/02161—Number of inputs available containing the signal or the noise to be suppressed
    • G10L2021/02166—Microphone arrays; Beamforming

Definitions

  • the present disclosure relates generally to acoustic signal processing, and in
  • Acoustic signal processing is applicable today to improve the quality of sound signals such as from microphones.
  • many devices such as handsets operate in the presence of sources of echoes, e.g., loudspeakers.
  • signals from microphones may occur in a noisy environment, e.g., in a car or in the presence of other noise.
  • MMSE minimum mean squared error
  • FIG. 1 shows a simplified block diagram of a system embodiment of the invention.
  • FIG. 2 shows a simplified flow chart diagram of one method embodiment of the invention.
  • FIG. 3A shows a simplified block diagram of a time-frame of samples being
  • FIG. 3B shows a simplified block diagram of banding frequency bins to a plurality of frequency bands.
  • FIG. 3C shows a simplified block diagram of the application of calculated gains to bins of sampled input data.
  • FIG. 3D shows a simplified block diagram of a synthesis process of converting output bins to frames of output samples.
  • FIG. 3E is a simplified block diagram of an output stage that can be included in
  • FIG. 4 depicts a two-dimensional plot representation of a banding matrix for banding a set of transform bins in accordance with some embodiments of the invention.
  • FIG. 5 depicts example shapes of the bands in the frequency domain on both a linear and logarithmic scale. Also shown in FIG. 5 is the sum of example band filters in accordance with some embodiments of the invention.
  • FIG. 6 shows time domain filter representations for several filter bands of example embodiments of banding.
  • FIG. 7 shows a normalization gain for banding to a plurality of frequency bands in accordance with some embodiments of the invention.
  • FIG. 8A and FIG. 8B show two decompositions of the signal power (or other
  • frequency domain amplitude metric in a band eventually to an estimate of the desired signal power (or other frequency domain amplitude metric).
  • FIGS. 9A, 9B and 9C show the probability density functions over time of the ratio, phase, and coherence spatial features, respectively, for diffuse noise and a voice signal.
  • FIG. 10 shows a simplified block diagram of an embodiment of gain calculator 129 of FIG. 1 according to an embodiment of the present invention.
  • FIG. 11 shows a flowchart of the gain calculation step and the post-processing step of FIG. 2 for those embodiment that include post-processing, together with the optional step of calculating and incorporating an additional echo gain, in accordance with an embodiment of the present invention.
  • FIG. 12 shows a probability density function in the form of a scaled histogram of signal power in a given band for the case of noise signal and voice signal.
  • FIG. 13 shows the distribution of FIG. 12, together with four suppression gain
  • FIG. 14 shows the histograms of FIG. 12 together with a sigmoid gain curve and a modified sigmoid-like gain curve determined according to alternate embodiments of the invention.
  • FIG. 15 shows what happens to the probability density functions of FIG. 12 after applying the sigmoid-like gain curve and the modified sigmoid-like gain curve of FIG. 14.
  • FIG. 16 shows a simplified block diagram of one processing apparatus embodiment that includes a processing system that has one or more processors and a storage subsystem, the processing apparatus for processing a plurality of audio inputs and one or more reference signal inputs according to an embodiment of the invention. DESCRIPTION OF EXAMPLE EMBODIMENTS
  • Embodiments of the present invention include a method, a system or apparatus, a tangible computer-readable storage medium configured with instructions that when executed by at least one processor of a processing system, cause processing hardware to carry out a method, and logic that can be encoded in one or more computer-readable tangible media and configured when executed to carry out a method.
  • the method is to process a plurality of input signals, e.g., microphone signals to simultaneously suppress noise, out-of-location signals, and in some embodiments, echoes.
  • Embodiments of the invention process sampled data in frames of samples, frame-by- frame.
  • the term "instantaneous" in the context of such processing frame -by-frame means for the current frame.
  • Particular embodiments include a system comprising an input processor to accept a plurality of sampled input signals and form a mixed-down banded instantaneous frequency domain amplitude metric of the input signals for a plurality of frequency bands.
  • the input processor includes input transformers to transform to frequency bins, a downmixer, e.g., beamformer to form a mixed-down, e.g., beamformed signal, and a spectral banding element to form frequency bands.
  • the downmixing, e.g., beamforming is carried out prior to transforming, and in others, the transforming is prior to downmixing, e.g., beamforming.
  • One system embodiment includes a banded spatial feature estimator to estimate
  • banded spatial features from the plurality of sampled input signals e.g., after transforming, and in other embodiments, before transforming.
  • Versions of the system that include echo suppression include a reference signal input processor to accept one or more reference signals, a transformer and a spectral banding element to form a banded frequency domain amplitude metric representation of the one or more reference signals.
  • Such versions of the system include a predictor of a banded frequency domain amplitude metric representation of the echo based on adaptively determined filter coefficients.
  • a noise estimator determines an estimate of the banded spectral amplitude metric of the noise.
  • a voice-activity detector uses the banded spectral amplitude metric of the noise, an estimate of the banded spectral amplitude metric of the mixed-down signal determined by a signal spectral estimator, and previously predicted echo spectral content to ascertain whether there is voice or not.
  • the banded signal is a sufficiently accurate estimate of the banded spectral amplitude metric of the mixed-down signal, so that signal spectral estimator is not used.
  • the output of the VAD is used by an adaptive filter updater to determine whether or not to update the filter coefficients, the updating based on the estimates of the banded spectral amplitude metric of the mixed-down signal and of the noise, and the previously predicted echo spectral content.
  • the system further includes a gain calculator to calculate suppression probability indicators, e.g., as gains including, in one embodiment, an out-of-location signal probability indicator, e.g., out-of-location gain determined using two or more of the spatial features, and a noise suppression probability indicator, e.g., noise suppression gain determined using an estimate of noise spectral content.
  • the estimate of noise spectral content is a spatially-selective estimate of noise spectral content.
  • the noise suppression probability indicator e.g., suppression gain includes echo suppression.
  • the gain calculator further is to combine the raw suppression probability indicators, e.g., suppression gains to a first combined gain for each band.
  • the gain calculator further is to carry out post-processing on the first combined gains of the bands to generate a post-processed gain for each band.
  • the post-processing includes depending on the version, one or more of: ensuring minimum gain, in some embodiments in a band dependent manner; in some embodiments ensuring there are no outlier or isolated gains by carrying out median filtering of the combined gain; and in some embodiments ensuring smoothness by carrying out time smoothing and, in some embodiments, band-to-band smoothing.
  • such post-processing includes spatially-selective voice activity detecting using two or more of the spatial features to generate a signal classification, such that the postprocessing is according to the signal classification.
  • the gain calculator further calculates an additional echo
  • suppression probability indicator e.g., an echo suppression gain. In one embodiment this is combined with the other gains (prior to post-processing in embodiments that include postprocessing) to form the first combined gain, which is a final gain. In another embodiment, the additional echo suppression probability indicator, e.g., suppression gain is combined, with the results of post-processing in embodiments that include post-processing, otherwise with the first combined gain to generate the final gain.
  • the system further includes a noise suppressor that interpolates the final gain to
  • the system further includes one or both of: a) an output synthesizer and transformer to generate output samples in the time domain, and b) output remapping to generate output frequency bins suitable for use by a subsequent codec or processing stage.
  • Particular embodiments include a system comprising means for accepting a plurality of sampled input signals and forming a mixed-down banded instantaneous frequency domain amplitude metric of the input signals for a plurality of frequency bands.
  • the means for accepting and forming includes means for transforming to frequency bins, means for downmixing, e.g., for beamforming to form a mixed-down, e.g., beamformed signal, and means for banding to form frequency bands.
  • the beamforming is carried out prior to transforming, and in other embodiments, the
  • One system embodiment includes means for determining banded spatial features from the plurality of sampled input signals.
  • Some system embodiments that include echo suppression include means for accepting one or more reference signals and for forming a banded frequency domain amplitude metric representation of the one or more reference signals, and means for predicting a banded frequency domain amplitude metric representation of the echo.
  • the means for predicting includes means for adaptively determining echo filter coefficients coupled to means for determining an estimate of the banded spectral amplitude metric of the noise, means for voice-activity detecting (VAD) using the estimate of the banded spectral amplitude metric of the mixed-down signal, and means for updating the filter coefficients based on the estimates of the banded spectral amplitude metric of the mixed-down signal and of the noise, and the previously predicted echo spectral content.
  • VAD voice-activity detecting
  • the means for updating updates according to the output of the means for voice activity detecting.
  • One system embodiment further includes means for calculating suppression probability indicators, e.g., suppression gains including an out-of-location signal gain determined using two or more of the spatial features, and a noise suppression probability indicator, e.g., noise suppression gain determined using an estimate noise spectral content.
  • the estimate of noise spectral content is a spatially-selective estimate of noise spectral content.
  • the noise suppression probability indicator e.g., suppression gain includes echo suppression.
  • the calculating by the means for calculating includes combining the raw suppression probability indicators, e.g., suppression gains to form a first combined gain for each band.
  • the means for calculating further includes means for carrying out post-processing on the first combined gains of the bands to generate a post- processed gain for each band.
  • the post-processing includes depending on the embodiment, one or more of: ensuring minimum gain, in some embodiments in a band dependent manner; in some embodiments ensuring there are no outlier or isolated gains by carrying out median filtering of the combined gain; and in some embodiments ensuring smoothness by carrying out time smoothing and, in some embodiments, band-to-band smoothing.
  • the means for post-processing includes means for spatially-selective voice activity detecting using two or more of the spatial features to generate a signal classification, such that the post-processing is according to the signal classification.
  • the means for calculating includes means for calculating an additional echo suppression probability indicator, e.g., suppression gain. This is combined in some embodiments with gain(s) (prior to post-processing in embodiments that include postprocessing) to form the first combined gain, with the post-processing first combined gain forming a final gain, and in other embodiments, the additional echo suppression probability indicator, e.g., suppression gain is combined with the results of post-processing in embodiments that include post-processing, otherwise with the first combined gain to generate a final gain.
  • an additional echo suppression probability indicator e.g., suppression gain
  • One system embodiment further includes means for interpolating the final gain to bin gains and for applying the final bin gains to carry out suppression on the bin data of the mixed-down signal to form suppressed signal data.
  • One system embodiment further includes means for applying one or both of: a) output synthesis and transforming to generate output samples, and b) output remapping to generate output frequency bins.
  • Particular embodiments include a processing apparatus comprising a processing system and configured to suppress undesired signals including noise and out-of-location signals, the processing apparatus configured to: accept a plurality of sampled input signals and form a mixed-down banded instantaneous frequency domain amplitude metric of the input signals for a plurality of frequency bands, the forming including transforming into complex- valued frequency domain values for a set of frequency bins.
  • the processing apparatus is further configured to determine banded spatial features from the plurality of sampled input signals; to calculate a first set of suppression probability indicators, including an out-of-location suppression probability indicator determined using two or more of the spatial features, and a noise suppression probability indicator for each band determined using an estimate of noise spectral content; to combine the first set of probability indicators to determine a first combined gain for each band; and to apply an interpolated final gain determined from the first combined gain to carry out suppression on bin data of the mixed- down signal to form suppressed signal data.
  • the estimate of noise spectral content is a spatially-selective estimate of noise spectral content determined using two or more of the spatial features.
  • Particular embodiments include a method of operating a processing apparatus to
  • the method comprises: accepting in the processing apparatus a plurality of sampled input signals, and forming a mixed-down banded instantaneous frequency domain amplitude metric of the input signals for a plurality of frequency bands, the forming including downmixing, e.g., transforming into complex- valued frequency domain values for a set of frequency bins.
  • the forming includes transforming the input signals to frequency bins, downmixing, e.g., beamforming the frequency data, and banding.
  • the downmixing can be before transforming, so that a single mixed-down signal is transformed.
  • the method includes determining banded spatial features from the plurality of
  • the method includes accepting one or more reference signals and forming a banded frequency domain amplitude metric representation of the one or more reference signals.
  • the representation in one embodiment is the sum.
  • the method includes predicting a banded frequency domain amplitude metric representation of the echo using adaptively updated echo filter coefficients, the coefficients updated using an estimate of the banded spectral amplitude metric of the noise, previously predicted echo spectral content, and an estimate of the banded spectral amplitude metric of the mixed-down signal.
  • the estimate of the banded spectral amplitude metric of the mixed-down signal is in one embodiment the mixed-down banded instantaneous frequency domain amplitude metric of the input signals, while in other embodiments, signal spectral estimation is used.
  • the control of the update of the prediction filter in one embodiment further includes voice-activity detecting— VAD— using the estimate of the banded spectral amplitude metric of the mixed- down signal, the estimate of banded spectral amplitude metric of noise, and the previously predicted echo spectral content. The results of voice-activity detecting determine whether there is updating of the filter coefficients.
  • the updating of the filter coefficients is based on the estimates of the banded spectral amplitude metric of the mixed-down signal and of the noise, and the previously predicted echo spectral content.
  • the method includes calculating raw suppression probability indicators, e.g.,
  • suppression gains including an out-of-location signal gain determined using two or more of the spatial features and a noise suppression probability indicator, e.g., as a noise suppression gain determined using an estimate of noise spectral content, and combining the raw suppression probability indicators, e.g., suppression gains to determine a first combined gain for each band.
  • the estimate of noise spectral content is a spatially- selective estimate of noise spectral content.
  • the noise suppression probability indicator e.g., suppression gain in some embodiments includes suppression of echoes, and its calculating also uses the predicted echo spectral content.
  • the method further includes carrying out spatially-selective voice activity detection determined using two or more of the spatial features to generate a signal classification, e.g., whether the input audio signal is voice or not.
  • a signal classification e.g., whether the input audio signal is voice or not.
  • wind detection is used, such that the signal classification further includes whether the input audio signal is wind or not.
  • Some embodiments of the method further include carrying out post-processing on the first combined gains of the bands to generate a post-processed gain for each band.
  • the postprocessing includes in some embodiments one or more of: ensuring minimum gain, e.g., in a band dependent manner, ensuring there are no isolated or outlier gains by carrying out median filtering of the combined gain, and ensuring smoothness by carrying out time and/or band-to-band smoothing.
  • the post-processing is according to the signal classification.
  • the method includes
  • an additional echo suppression probability indicator e.g., suppression gain.
  • the additional echo suppression gain is combined with the other raw suppression gains to form the first combined gain, and (post-processed if post-processing is included) first combined gain forms a final gain for each band.
  • the additional echo suppression gain is combined with the (post-processed if post-processing is included) first combined gain to generate a final gain for each band.
  • the method includes interpolating the final gain to produce final bin gains, and
  • Particular embodiments include a method of operating a processing apparatus to suppress undesired signals, the undesired signals including noise.
  • Particular embodiments also include a processing apparatus including a processing system, with the processing apparatus configured to carry out the method. The method comprises: accepting in the processing apparatus at least one sampled input signal; and forming a banded instantaneous frequency domain amplitude metric of the at least one input signal for a plurality of frequency bands, the forming including transforming into complex- valued frequency domain values for a set of frequency bins.
  • the method further comprises calculating a first set of one or more suppression probability indicators, including a noise suppression probability indicator determined using an estimate of noise spectral content; combining the first set of probability indicators to determine a first combined gain for each band; and applying an interpolated final gain determined from the first combined gain to carry out suppression on bin data of the at least one input signal to form suppressed signal data.
  • the noise suppression probability indicator for each frequency band is expressible as noise suppression gain function of the banded instantaneous amplitude metric for the band. For each frequency band, a first range of values of banded instantaneous amplitude metric values is expected for noise, and a second range of values of banded instantaneous amplitude metric values is expected for a desired input.
  • the noise suppression gain functions for the frequency bands are configured to: have a respective minimum value; have a relatively constant value or a relatively small negative gradient in the first range; have a relatively constant gain in the second range; and have a smooth transition from the first range to the second range.
  • Particular embodiments include a method of operating a processing apparatus to suppress undesired signals.
  • the method comprises: accepting in the processing apparatus at least one sampled input signal; forming a banded instantaneous frequency domain amplitude metric of the at least one input signal for a plurality of frequency bands, the forming including transforming into complex-valued frequency domain values for a set of frequency bins; calculating a first set of one or more suppression probability indicators, including a noise suppression probability indicator determined using an estimate of noise spectral content; and combining the first set of probability indicators to determine a first combined gain for each band.
  • Some embodiments of the method further comprise carrying out postprocessing on the first combined gains of the bands to generate a post-processed gain for each band, the post-processing including ensuring minimum gains for each band; and applying an interpolated final gain determined from the post-processed gain to carry out suppression on bin data of the at least one input signal to form suppressed signal data.
  • the post-processing includes one or more of: carrying out median filtering of gains; carrying out band-to-band smoothing of gains, and carrying out time smoothing of gains.
  • Particular embodiments include a method of operating a processing apparatus to process at least one sampled input signal, the method comprising: accepting in the processing apparatus at least one sampled input signal and forming a banded instantaneous frequency domain amplitude metric of the at least one input signal for a plurality of frequency bands, the forming including transforming into complex-valued frequency domain values for a set of frequency bins and banding to a plurality of frequency bands.
  • the method further includes calculating a gain for each band in order to achieve noise reduction and/or, in the case that the banding is perceptual banding, one or more of perceptual domain-based leveling, perceptual domain-based dynamic range control, and perceptual domain-based dynamic equalization.
  • the method further comprises carrying out postprocessing on the gains of the bands to generate a post-processed gain for each band; the post-processing including median filtering of the gains of the bands, and applying an interpolated final gain determined from the (post-processed if post-processing is included) gain to carry out noise reduction and/or, in the case that the banding is perceptual banding, one or more of perceptual domain-based leveling, perceptual domain-based dynamic range control, and perceptual domain-based dynamic equalization on bin data to form processed signal data.
  • Some versions of the method further comprise carrying out at least one of voice activity detecting and wind activity detecting to a signal classification, wherein the median filtering depends on the signal classification.
  • Particular embodiments include a method of operating a processing apparatus to suppress undesired signals, the method comprising: accepting in the processing apparatus a plurality of sampled input signals; and forming a mixed-down banded instantaneous frequency domain amplitude metric of the input signals for a plurality of frequency bands, the forming including transforming into complex-valued frequency domain values for a set of frequency bins.
  • the method further comprises determining banded spatial features from the plurality of sampled input signals; calculating a first set of suppression probability indicators, including an out-of-location suppression probability indicator determined using two or more of the spatial features, and a noise suppression probability indicator determined using an estimate of noise spectral content; combining the first set of probability indicators to determine a first combined gain for each band.
  • the first combined gain after post-processing if post-processing is included, forms a final gain for each band. ; and applying an interpolated final gain determined from the first combined gain. Interpolating the final gain produces final bin gains to apply to bin data of the mixed-down signal to form suppressed signal data.
  • the estimate of noise spectral content is a spatially-selective estimate of noise spectral content determined using two or more of the spatial features.
  • the estimate noise spectral content is determined by a leaky minimum follower with a tracking rate defined by at least one minimum follower leak rate parameter.
  • the at least one leak rate parameter of the leaky minimum follower are controlled by the probability of voice being present as determined by voice activity detecting.
  • Particular embodiments include a method of operating a processing apparatus to suppress undesired signals, the method comprising: accepting in the processing apparatus a plurality of sampled input signals; forming a mixed-down banded instantaneous frequency domain amplitude metric of the input signals for a plurality of frequency bands, the forming including transforming into complex-valued frequency domain values for a set of frequency bins; and determining banded spatial features from the plurality of sampled input signals.
  • the method further comprises calculating a first set of suppression probability indicators, including an out-of-location suppression probability indicator determined using two or more of the spatial features, and a noise suppression probability indicator determined using an estimate of noise spectral content; accepting in the processing apparatus one or more reference signals; forming a banded frequency domain amplitude metric representation of the one or more reference signals; and predicting a banded frequency domain amplitude metric representation of an echo using adaptively determined echo filter coefficients.
  • the method further includes determining a plurality of indications of voice activity from the mixed-down banded instantaneous frequency domain amplitude metric using respective instantiations of a universal voice activity detection method, the universal voice activity detection method controlled by a set of parameters and using: an estimate of noise spectral content, the banded frequency domain amplitude metric representation of the echo, and the banded spatial features, the set of parameters including whether the estimate of noise spectral content is spatially selective or not, which indication of voice activity an instantiation determines being controlled by a selection of the parameters, voice activity.
  • the method further comprises combining the first set of probability indicators to determine a first combined gain for each band; and applying an interpolated final gain determined from the gain (post-processed, if post-processing is included) to carry out suppression on bin data of the mixed-down signal to form suppressed signal data.
  • Different instantiations of the universal voice activity detection method are applied in different steps of the method.
  • the estimate of noise spectral content is a spatially-selective estimate of noise spectral content determined using two or more of the spatial features.
  • Particular embodiments include a tangible computer-readable storage medium
  • Particular embodiments include logic that can be encoded in one or more computer- readable tangible media to carry out a method as described herein.
  • Particular embodiments may provide all, some, or none of these aspects, features, or advantages. Particular embodiments may provide one or more other aspects, features, or advantages, one or more of which may be readily apparent to a person skilled in the art from the figures, descriptions, and claims herein. Particular example embodiments
  • Described herein is a method of processing: (a) a plurality of input signals, e.g., signals from a plurality of spatially separated microphones; and, for echo suppression, (b) one or more reference signals, e.g., signals from or to be rendered by one or more loudspeakers and that can cause echoes.
  • a source of sound e.g., a human who is a source of human voice for the array of microphones.
  • the method processes the input signals and one or more reference signals to carry out in an integrated manner simultaneous noise suppression, echo suppression, and out-of-location signal suppression.
  • Also described herein is a system accepting the plurality of input signals and the one or more reference signals to process the input signals and one or more reference signals to carry out in an integrated manner simultaneous noise suppression, echo suppression, and out-of-location signal suppression.
  • at least one storage medium on which are coded instructions that when executed by one or more processors of a processing system, cause processing a plurality of input signals, e.g., microphone signals and one or more reference signals, e.g., for or from one or more loudspeakers to carry out in an integrated manner simultaneous noise suppression, echo suppression, and out-of-location signal suppression.
  • Embodiments of the invention are described in terms of determining and applying a set of suppression probability indicators, expressed, e.g., as suppression gains for each of a plurality of spectral bands, applied to spectral values of signals at a number of frequency bands.
  • the spectral values represent spectral content.
  • the spectral content is in terms of the power spectrum.
  • the invention is not limited to processing power spectral values. Rather, any spectral amplitude dependent metric can be used. For example, if the amplitude spectrum is used directly, such spectral content is sometimes referred to as spectral envelope. Thus, often, rather than using the phrase "power spectrum,” the phrase “power spectrum (or other amplitude metric spectrum)" is used in the description.
  • B The number of spectral values, also called the number of bands.
  • the B bands are at frequencies whose spacing is monotonically non- decreasing. At least 90% of the frequency bands include contribution from more than one frequency bin, and in a preferred embodiment, each frequency band includes contribution from two or more frequency bins.
  • the bands are monotonically increasing in a log-like manner. In some particular embodiments, they are on a psycho-acoustic scale, that is, the frequency bands are spaced with a scaling related to psycho-acoustic critical spacing, such banding called perceptually-banding" herein b: The band number from 1 to B. fc (b) : The center frequency of band b.
  • N The number of frequency bins after transforming to the frequency domain.
  • M The number of samples in a frame, e.g., the number of samples being windowed by a suitable window.
  • T The time interval of the sound being sampled by a frame of M samples.
  • Q The sampling frequency for the M samples of a frame.
  • P The number of input signals, e.g., microphone input signals.
  • Q The number of reference inputs.
  • the instantaneous (banded) spectral content e.g., instantaneous spectral power (or other frequency domain amplitude metric) in the mixed-down, e.g., beamformed signal (combined with noise and echo) of the most recent ⁇ -long frame (the current frame) in frequency band b. This is determined, e.g., by banding into frequency bands the mixed-down, e.g., beamformed transformed signal bins.
  • the N frequency bins of the reference input of the most recent ⁇ -long frame (the current frame) of M samples obtained e.g., by transforming into frequency bands a signal representative of the one or more reference inputs.
  • the reference input instantaneous spectral content, e.g., instantaneous power (or other frequency domain amplitude metric) of the most recent ⁇ -long frame (the current frame) in frequency band b. This is determined, e.g., by transforming and banding into frequency bands a signal representative of the one or more reference inputs.
  • Eb' The predicted echo spectral content, e.g., power spectrum (or other amplitude metric spectrum) in frequency band b.
  • the signal estimated spectral content e.g., power spectrum (or other amplitude metric spectrum) of the most recent frame (the current frame) in frequency band b, determined from the instantaneous banded power 3 ⁇ 4' .
  • 3 ⁇ 4 may be a sufficiently good estimate of i3 ⁇ 4' .
  • the noise estimate spectral content e.g., power spectrum (or other amplitude metric spectrum) in frequency band b. This is used, e.g., for voice activity detection and for updating filter coefficients for the adaptive prediction of the echo spectral content.
  • FIG. 1 shows a block diagram of an embodiment of a system 100 that accepts a
  • the signals 101 and 102 are in the form of sample values. In some embodiments, a number of one or more denoted P of signal inputs 101, e.g., microphone inputs from microphones (not shown) at different respective spatial locations, the input signals denoted MIC 1, ..., MIC P, and a number, denoted Q of reference inputs 102, denoted REF 1, REF ⁇ 2, e.g., ⁇ 2 inputs 102 to be rendered on Q loudspeakers, or signals obtained from Q loudspeakers.
  • the signals 101 and 102 are in the form of sample values. In some embodiments, MIC 1, ..., MIC P, and a number, denoted Q of reference inputs 102, denoted REF 1, REF ⁇ 2, e.g., ⁇ 2 inputs 102 to be rendered on Q loudspeakers, or signals obtained from Q loudspeakers.
  • the signals 101 and 102 are in the form of sample values. In some embodiments, MIC 1, ..., MIC
  • P>2 so that there are at least two signal inputs, e.g., microphone inputs.
  • Q> ⁇ the system 100 shown in FIG. 1 carries out in an integrated manner simultaneous noise suppression and out-of-location signal suppression, and in some embodiments also simultaneous echo suppression.
  • One such embodiment includes a system 100 comprising an input processor 103, 107, 109 to accept a plurality of sampled input signals and form a mixed-down banded instantaneous frequency domain amplitude metric 110 of the input signals 101 for a plurality B of frequency bands.
  • the beamforming is carried out prior to transforming, and in others, as shown in FIG. 1, the transforming is prior to downmixing, e.g., beamforming.
  • One system embodiment includes a banded spatial feature estimator 105 to estimate banded spatial features 106 from the plurality of sampled input signals, e.g., after transforming, and in other embodiments, before transforming.
  • Versions of system 100 that include echo suppression include a reference signal input processor 111 to accept one or more reference signals, a transformer 113 and a spectral banding element 115 to form a banded frequency domain amplitude metric representation 116 of the one or more reference signals.
  • Such versions of system 100 include a predictor 117 of a banded frequency domain amplitude metric representation of the echo 118 based on adaptively determined filter coefficients.
  • a noise estimator 123 determines an estimate of the banded spectral amplitude metric of the noise 124.
  • a voice- activity detector (VAD) 124 uses the banded spectral amplitude metric of the noise 124, an estimate of the banded spectral amplitude metric of the mixed-down signal 122 determined by a signal spectral estimator 121, and previously predicted echo spectral content 118 to produce a voice detection output.
  • the banded signal 110 is a sufficiently accurate estimate of the banded spectral amplitude metric of the mixed- down signal 122, so that signal spectral estimator 121 is not used.
  • the results of the VAD 125 are used by an adaptive filter updater 127 to determine whether to update the filter coefficients 128 based on the estimates of the banded spectral amplitude metric of the mixed- down signal 122 (or 110) and of the noise 124, and the previously predicted echo spectral content 118.
  • System 100 further includes a gain calculator 129 to calculate suppression probability indicators, e.g., as gains including, in one embodiment, an out-of-location signal probability indicator, e.g., gain determined using two or more of the spatial features 106, and a noise suppression probability indicator, e.g., gain determined using spatially-selective noise spectral content.
  • the noise suppression gain includes echo suppression.
  • the gain calculator 129 further is to combine the raw suppression gains to a first combined gain for each band.
  • gain calculator 129 further is to carry out post-processing on the first combined gains of the bands to generate a post-processed gain 130 for each band.
  • the post-processing includes depending on the embodiment, one or more of: ensuring minimum gain, in some embodiments in a band dependent manner; in some embodiments ensuring there are no outlier or isolated gains by carrying out median filtering of the combined gain; and in some embodiments ensuring smoothness by carrying out time smoothing and, in some embodiments, band-to-band smoothing.
  • the post-processing includes spatially-selective voice activity detecting using two or more of the spatial features 106 to generate a signal classification, such that the post-processing is according to the signal classification.
  • the gain calculator 129 further calculates an additional echo suppression gain. In one embodiment this is combined with the other gains (prior to post- processing, if post-processing is included) to form the first combined gain. In another embodiment, the additional echo suppression gain is combined with the first combined gain (after post-processing, if post-processing is included) to generate a final gain for each band.
  • System 100 further includes a noise suppressor 131 to apply the gain 130 (after postprocessing, if post-processing is included) to carry out suppression on the bin data of the mixed-down signal to form suppressed signal data 132.
  • System 100 further includes in 133 one or both of: a) an output synthesizer and transformer to generate output samples, and b) output remapping to generate output frequency bins.
  • System embodiments of the invention include a system comprising: means for
  • the means for accepting and forming includes means 103 for transforming to frequency bins, means 107 for beamforming to form a mixed-down, e.g., beamformed signal, and means for banding (109) to form frequency bands.
  • the beamforming is carried out prior to transforming, and in others, the transforming is prior to downmixing, e.g., beamforming.
  • One system embodiment includes means for determining 105 banded spatial features 106 from the plurality of sampled input signals.
  • the system embodiments that include echo suppression include means for accepting 213 one or more reference signals and for forming 215, 217 a banded frequency domain amplitude metric representation 116 of the one or more reference signals, and means for predicting 117, 123, 125, 127 a banded frequency domain amplitude metric representation of the echo 118.
  • the means for predicting 117, 123, 125, 127 includes means for adaptively determining 125, 127 echo filter coefficients 128 coupled to means for determining 123 an estimate of the banded spectral amplitude metric of the noise 124, means for voice-activity detecting (VAD) using the estimate of the banded spectral amplitude metric of the mixed-down signal 122, and means for updating 127 the filter coefficients 128.
  • VAD voice-activity detecting
  • the output of the VAD is coupled to means for updating and determined if the means for updating updates the filter coefficients.
  • One system embodiment further includes means for calculating 129 suppression gains including an out-of-location signal gain determined using two or more of the spatial features 106, and a noise suppression gain determined using spatially-selective noise spectral content.
  • the noise suppression gain includes echo suppression.
  • the calculating of the means for calculating 129 includes combining the raw suppression gains to a first combined gain for each band.
  • the means for calculating 129 further includes means for carrying out post-processing on the first combined gains of the bands to generate a post- processed gain 130 for each band.
  • the post-processing includes in some embodiments one or more of ensuring minimum gain, e.g., in a band dependent manner, ensuring there are no isolated gains by carrying out median filtering of the combined gain, and ensuring smoothness by carrying out time and/or band-to-band smoothing.
  • the means for post-processing includes means for spatially-selective voice activity detecting using two or more of the spatial features 106 to generate a signal classification, such that the post-processing is according to the signal classification.
  • the means for calculating 129 includes means for calculating an additional echo suppression gain. This is combined in some embodiments with gain(s) (prior to post-processing, if post-processing is included) to form the first combined gains of the bands to be used as a final gain for each band, and in other embodiments the additional echo suppression gain in each band is combined with the first combined gains (post- processed, if post-processing is included) to generate a final gain for each band.
  • One system embodiment further includes means 131 for interpolating the final gains to final bin gains and applying the final bin gains to carry out suppression on the bin data of the mixed-down signal to form suppressed signal data 132.
  • One system embodiment further includes means 133 for applying one or both of: a) output synthesis and transforming to generate output samples 135, and b) output remapping to generate output frequency bins 135 (note the same reference numeral is used for both an output sample generator, and an output frequency bin generator).
  • method 200 shows a flowchart of a method 200 of operating a processing apparatus 100 to suppress noise and out-of-location signals and in some embodiments echo in a number denoted P of signal inputs 101, e.g., microphone inputs from microphones at different respective spatial locations, the input signals denoted MIC 1, ..., MIC P.
  • method 200 includes processing a number, denoted Q of reference inputs 102, denoted REF 1, ..., REF Q, e.g., Q inputs to be rendered on Q loudspeakers, or signals obtained from Q loudspeakers.
  • the signals are in the form of sample values.
  • the system carries out, in an integrated manner, simultaneous noise suppression, out-of-location signal suppression, and, in some embodiments, echo suppression.
  • method 200 comprises: accepting 201 in the processing
  • the apparatus a plurality of sampled input signals 101, and forming 203, 207, 209 a mixed-down banded instantaneous frequency domain amplitude metric 110 of the input signals 101 for a plurality of frequency bands, the forming including transforming 203 into complex-valued frequency domain values for a set of frequency bins.
  • the forming includes in 203 transforming the input signals to frequency bins, downmixing, e.g., beamforming the frequency data, and in 207 banding.
  • the downmixing can be before transforming, so that a single mixed-down signal is transformed.
  • the system may make use of an estimate of the banded echo reference, or a similar representation of the frequency domain spectrum of the echo reference provided by another processing component or source within the realized system.
  • the method includes determining in 205 banded spatial features 106 from the
  • the method includes accepting 213 one or more reference signals and forming in 215 and 217 a banded frequency domain amplitude metric representation 116 of the one or more reference signals.
  • the representation in one embodiment is the sum.
  • the method includes predicting in 221 a banded frequency domain amplitude metric representation of the echo 118 using adaptively determined echo filter coefficients 128.
  • the predicting in one embodiment further includes voice-activity detecting— VAD— using the estimate of the banded spectral amplitude metric of the mixed-down signal 122, the estimate of banded spectral amplitude metric of noise 124, and the previously predicted echo spectral content 118.
  • the coefficients 128 are undated or not according to the results of voice-activity detecting. Updating uses an estimate of the banded spectral amplitude metric of the noise 124, previously predicted echo spectral content 118, and an estimate of the banded spectral amplitude metric of the mixed-down signal 122.
  • the estimate of the banded spectral amplitude metric of the mixed-down signal is in one embodiment the mixed-down banded instantaneous frequency domain amplitude metric 110 of the input signals, while in other embodiments, signal spectral estimation is used.
  • the method 200 includes: a) calculating in 223 raw
  • the method 200 further includes carrying out in spatially- selective voice activity detection determined using two or more of the spatial features 106 to generate a signal classification, e.g., whether voice or not.
  • a signal classification e.g., whether voice or not.
  • wind detection is used such that the signal classification further includes whether the signal is wind or not.
  • the method 200 further includes carrying out post-processing on the first combined gains of the bands to generate a post-processed gain 130 for each band.
  • the post-processing includes in some embodiments one or more of: ensuring minimum gain, e.g., in a band dependent manner, ensuring there are no isolated gains by carrying out median filtering of the combined gain, and ensuring smoothness by carrying out time and/or band-to- band smoothing.
  • the post-processing is according to the signal classification.
  • the method includes
  • the method includes applying in 227 the final gain, including interpolating the gain for bin data to carry out suppression on the bin data of the mixed-down signal to form suppressed signal data 132. And apply in 229 one or both of a) output synthesis and transforming to generate output samples, and b) output remapping to generate output frequency bins.
  • FIG. 1 that includes all aspects of suppression, including simultaneous echo, noise, and out-of- spatial location suppression, or presented as a computer-readable storage medium that includes instructions that when executed by one or more processors of a processing system (see FIG. 16 and description thereof), cause a processing apparatus that includes the processing system to carry out the method such as that of FIG. 2, note that the example embodiments also provide a scalable solution for simpler applications and situations.
  • One embodiment includes simultaneous noise suppression, echo suppression and out- of-spatial location suppression, while another embodiment includes simultaneous noise suppression and out-of-spatial location suppression.
  • Much of the description herein assumes simultaneous noise suppression, echo suppression and out-of-location signal suppression, and how to modify any embodiment to not include echo suppression would be clear to one skilled in the art.
  • the reference signals and input signals
  • the Q reference signals represent a set of audio signals that relate to the potential echo at the microphone array.
  • the microphone array may be that of a headset, personal mobile device or fixed microphone array.
  • the references may correspond to signals being used to drive one or several speakers on the headset or personal mobile device, or one or more speakers used in a speaker array or surround sound configuration, or the loudspeakers on a portable device such as a laptop computer or tablet. It is noted that the application is not limited to these scenarios, however the nature of the approach is best suited to an environment where the response from each reference to the microphone array center is similar in gain and delay.
  • the reference signals may also represent a signal representation prior to the actual speaker feeds, for example a raw audio stream prior to it being rendered and sent to a multichannel speaker output.
  • the proposed approach offers a solution for robust echo control which also allows for moderate spatial and temporal variation in the echo path, including being robust to sampling offsets, discontinuities and timing drift.
  • the reference inputs may represent the output speaker feeds that are creating the potential echo, or alternately the sources that will be used to create the speaker outputs after appropriate rendering. The system will work well for either case, however in some embodiments, the use of the initial independent and likely uncorrected sources prior to rendering are preferred.
  • the adaptive framework presented in this invention is able to manage the variation and complexity of the multi channel echo source.
  • the use of the component audio sources rather than the rendered speaker feeds can be beneficial in avoiding issues in the combination of the echo reference due to signal correlations.
  • the combination of the echo reference and robustness for the multichannel echo suppression is discussed further later in the disclosure.
  • the output of the system is a single signal representing the separated voice or signal of interest after the removal of noise, echo and sound components not originating from the desired position.
  • the output of the system is a set of remapped frequency components representing the separated voice or signal of interest after the removal of noise, echo and sound components not originating from the desired position. These frequency components are, e.g., in a form usable by a subsequent compression (coding) method or additional processing component.
  • Each of the processing of system 100 and the method 200 is carried out in a frame- based manner (also called block-based manner) on a frame of M input samples (also called a block of M input samples) at each processing time instant.
  • the P inputs e.g., microphone inputs are transformed by one or more time-to-frequency transformers 103 independently to produce a set of P frequency domain representations.
  • the transform to the frequency domain representation will typically have a set of N linearly spaced frequency bins each having a single complex value at each processing time instant. It is noted that generally N>M such that at each time instant, M new audio data samples are processed to create N complex- valued frequency domain representation data points.
  • the increased data in the complex- valued frequency domain representation allows for a degree of analysis and processing of the audio signal suited to the noise, echo and spatial selectivity algorithm to achieve reasonable phase estimation.
  • the Q reference inputs are combined using a simple time domain sum. This creates a single reference signal of M real-valued samples at each processing instant. It has been found by the inventor(s) that the system is able to achieve suppression for a multi-channel echo by using only a single combined reference. While the invention does not depend on any reasoning of why the results are achieved, it is believed that using only a single combined reference works, we believe, as a result of the inherent robustness of using the banded amplitude metric representation of the echo, noise and signal within the suppression framework, and the broader time resolution offered from the time-frame-based processing. This approach allows a certain timing and gain uncertainty or margin of error.
  • the Q reference inputs are combined, e.g., using summation in the time domain to create a single reference signal to be used for the echo control. In some embodiments, this summation may occur after the transform or at the banding stage where the power spectra (or other amplitude metric spectra) of the Q reference signals may be combined. Combining the signals in the power domain has the advantage of avoiding the effects of destructive (cancellation) or constructive combination of correlated content across the ⁇ 2 signals. Such 'in phase' or exact phase aligned combination of the reference signals is unlikely to occur extensively and consistently across time and/or frequency at the microphones due to the inherent complexities of the expected acoustic echo paths.
  • the direct combination approach can create deviations in the single channel reference power estimate and its ability to be used as an echo predictor. In practice, this is not found to be a significant problem for typical multi channel content.
  • the single channel time domain summation offers effective performance at very low complexity. Where a large amount of correlated content is expected between the channels, and the probability is reasonable that there may be opposing phase and time aligned content, the potential for loss of echo control performance can be reduced by using a de-correlating filter on one or more of the reference channels.
  • a de-correlating filter on one or more of the reference channels.
  • a de-correlating filter on one or more of the reference channels.
  • a time delay A 2-5ms time delay is suggested for such embodiments of the invention.
  • Another example is a bulk phase shift such as a Hilbert transform or 90-degree phase shift.
  • Embodiments of the invention process the data frame-by-frame, with each
  • FIGS. 3A-3E show some details of some of the elements of embodiments of the invention.
  • FIG. 3 A shows a frame (a block) of M input samples being placed in a buffer of length 2N with a set of 2N—M previous samples and being windowed according to a window function to generate 2N values which are transformed according to a transform, with an additional twist function as described below. This results in N complex- valued bins.
  • FIG. 3B shows the conversion of the N bins to a number B of frequency bands. The banding to B bands is described in more detail below.
  • One aspect of the invention is the determination of a set of B suppression gains for the B bands. The determination of the gains incorporates statistical spatial information, e.g., indicative of out-of-location signals.
  • FIG. 3C shows the interpolation of B gains to create a set of N gains which are then applied to N bins of input data.
  • Some embodiments of the invention include post-processing of raw-gains to ensure stability.
  • the post-processing is controlled based on signal classification, e.g., a classification of the signal to according to one or more of (spatially selective) voice activity and wind activity.
  • the post-processing applied is selected according to signal activity classification.
  • the post-processing includes preventing the gains from falling below some pre- specified (frequency-band-dependent) minimum point, the manner of prevention dependent on the activity classification, how musical noise due to one or more isolated gain values can be effectively eliminated in a manner dependent on the activity classification, and how the gains may be smoothed, with the type and amount of smoothing dependent on the activity classification.
  • FIG. 3D describes the synthesis process of converting the ⁇ output bins to a frame of M output samples, and typically involves inverse transforming and windowed overlap-add operations.
  • FIG. 3E is an optional output stage which can reformat the N complex- valued bins from FIG. 3C to suit the transform needs of subsequent processing (such as an audio codec) thus saving processing time and reducing signal latency.
  • subsequent processing such as an audio codec
  • FIG. 3D the processing of FIG. 3D is not used, as the output is to be encoded in some manner. In such cases, a remap operation as shown in FIG. 3E is applied.
  • DFT discrete finite length Fourier transform
  • FFT fast Fourier transform
  • a discrete finite length Fourier transform, such as implemented by the FFT is often referred to as a circulant transform due to the implicit assumption that the signal in the transform window is in some way periodic or repetitive.
  • Most general forms of circulant transforms can be represented by buffering, a window, a twist (real value to complex value transformation) and a DFT, e.g., FFT.
  • An optional complex twist after the DFT can be used to adjust the frequency domain representation to match specific transform definitions.
  • This class of transforms includes the modified DFT (MDFT), the short time Fourier transform (STFT) and with a longer window and wrapping, a conjugate quadrature mirror filter (CQMF).
  • MDFT modified DFT
  • STFT short time Fourier transform
  • CQMF conjugate quadrature mirror filter
  • MDCT Modified discrete cosine transform
  • MDST modified discrete sine transform
  • the additional complex twist of the frequency domain bins is used, however this does not change the underlying frequency resolution or processing ability of the transform and thus can be left until the end of the processing chain, and applied in the remapping if required.
  • x n represents the last 2N input samples with x N _ x representing the most recent sample
  • X n represents the N complex-valued frequency bins in increasing frequency order.
  • the inverse transform or synthesis of FIG. 3D is represented in the last two equation lines.
  • y n represents the 2N output samples that result from the individual inverse transform prior to overlapping, adding and discarding as appropriate for the designed windows. It should be noted, that this transform has an efficient implementation as a block multiply and FFT.
  • Tof M/fg a time interval
  • N 128,256 or 512.
  • the transform can be run more often or "overs ampled.”
  • the window functions u n and v n have an effect on the finer details of the transform frequency resolution and the transition and interpolation of activity between adjacent time frames of processed data. Since the transform is processed in an overlapping manner, the window functions control the nature of this overlap. It should be known to someone skilled in the art that there are many possibilities of window function related to this aspect of signal processing, each with different properties and trade-offs.
  • this window extends over the complete range of 2N samples.
  • STFT short term Fourier transform
  • FIG. 3A and FIG. 3D also known as prototype filters, can be of length greater or smaller than the examples given herein.
  • a smaller window can be represented in the general form suggested above with a set of zero coefficients (zero padding).
  • a longer window is typically implemented by applying the window and then folding the signal into the transform processing range of the 2N samples. It is known that the window design affects certain aspects of: frequency resolution, independence of the frequency domain bins, latency, and processing distortions.
  • the standard complex-valued fast Fourier transform can be used in implementing the transforms used herein, so that this complete transform has an efficient implementation using a set of complex block multiplication and a standard FFT. While not meant to be limiting, such that other embodiments can use other designs, this design facilitates porting of the transform or filterbank by taking advantage of any standard existing optimized FFT implementation for the target processor platform.
  • this design facilitates porting of the transform or filterbank by taking advantage of any standard existing optimized FFT implementation for the target processor platform.
  • M frame size and positioning
  • twists there are many families of transforms represented by variations to the input and output windows and the frame size and positioning (M) and twists. Provided the windows are not sub-optimal, the main characteristics are the frequency sampling resolution (N), the underlying frequency resolution (related to the width and shape of the input window) and the frame size or stride between transforms (M).
  • the window and complex twist may be different for each of the inputs, e.g., microphone inputs to effect appropriate time delay to be used in the mixing down, e.g., beamforming and in the positional inference.
  • the method can be made reasonably independent of the transform, provided the frame size (or stride) is known in order to update all processing time constants accordingly.
  • the N complex-valued bins for each of the P inputs are used directly to create a set of positional estimates of spatial probability of activity. This is shown in FIG. 1 as banded spatial feature estimator 105 and in FIG. 2 as step 205. The details and operation of element 105 and step 205 are described in more detail below after a discussion of the downmixing, e.g., by beamforming.
  • the ⁇ complex- valued bins for each of the P inputs are combined to make a single frequency domain channel, e.g., using a downmixer, e.g., a beamformer 107.
  • a downmixer e.g., a beamformer 107.
  • the downmixer is a beamformer 107 designed to achieve some spatial selectivity towards the desired position.
  • the beamformer 107 is a linear time invariant process, i.e., a passive beamformer defined in general by a set of complex- valued frequency-dependent gains for each input channel. Longer time extent filtering may be included to create a selective temporal and spatial beamformer.
  • Possible beamforming structures include a real- valued gain and combination of the P signals, for example in the case of two microphones this might be a simple summation or difference.
  • the term beamforming as used herein means mixing-down, and may include some spatial selectivity.
  • the beamformer 107 (and beamforming step 207) can include adaptive tracking of the spatial selectivity over time, in which case the beamformer gains
  • the tracking is sufficiently slow such that the time varying process beamformer 107 can be considered static for time periods of interest. Hence, for simplicity, and for analysis of the short-term system performance, it is sufficient to assume this component is time invariant.
  • the downmixer e.g., beamformer 107 and step 207 include using complex- valued frequency-dependent gains (mixing coefficients) derived for each processing bin.
  • Such a filter may be designed to achieve a certain directivity that is relatively constant or suitably controlled across different frequencies.
  • the downmixer, e.g., beamformer 107 will be designed or adapted to achieve an improvement in the signal to noise ratio of the desired signal, relative to that which would be achieved by any one microphone input signal.
  • beamforming is a well- studied problem and there are many techniques for achieving a suitable beamformer or linear microphone array process to create the mixed- down, e.g., beamformed signal out of beamformer 107 and step 207.
  • the beamforming 207 by beamformer 107 includes the nulling or cancellation of specific signals arriving from one or more known locations of sources undesired signal, such as echo, noise, or other undesired signal. While “nulling” suggest reducing to zero, in this description, “nulling” means reducing the sensitivity; those skilled in the art would understand that "perfect" nulling is not typically achievable in practice.
  • the linear process of the beamformer is only able to null a small number (P-l) of independently located sources.
  • This limitation of the linear beamformer is complemented by the more effective spatial suppression described later as a part of some embodiments of the present invention.
  • the location of spatial response of the microphone array to the expected dominant echo path may be known and relatively constant.
  • the source of the echo would be known as coming from the speaker(s).
  • the beamformer is designed to null, i.e., provide zero or low relative sensitivity to sound arriving from the known location of source(s) of undesired signal.
  • Embodiments of the present invention can be used in a system or method that
  • each of the beamformer 107 and beamforming 207 is time invariant.
  • beamformer 107 uses for beamformer 107 a passive beamformer 107 that determines the simple sum of the two input channels.
  • one embodiment of beamforming 207 includes introducing a relative delay and differencing of the two input signals from the microphones. This substantially approximates a hypercardioid microphone directionality pattern.
  • the designed mixing of the P microphone inputs to achieve a single intermediary signal has a preferential sensitivity for the desired source.
  • the downmixer e.g., the beamforming 207 of
  • beamformer 107 weights the sets of inputs (as frequency bins) by a set of complex valued weights.
  • the beamforming weights of beamformer 107 are determined according to maximum-ratio combining (MRC).
  • MRC maximum-ratio combining
  • the beamformer 107 uses weights determined using zero-forcing. Such methods are well known in the art.
  • each frequency bin contains a contribution from more than one or more frequency bins, with at least 90% of the bands having contributions from two or more bins, the number of bins non- decreasing with frequency such that higher frequency bands have contribution from more bins than lower frequency bands.
  • FIG. 3B shows the conversion of the N bins to a number B of frequency bands carried out by banding elements 109 and 115, and banding steps. 209 and 217.
  • One aspect of the invention is the determination of a set of B suppression gains for the B bands.
  • the determination of the gains incorporates statistical spatial information.
  • the raw frequency domain representation data is required for the intermediate signal, as this will be used in the signal synthesis to the time domain, the raw frequency domain coefficients of the echo reference are not required and can be discarded after calculating the power spectra (or other amplitude metric spectra).
  • the full set of P frequency domain representations of the microphone inputs is required to infer the spatial properties of the incident audio signal.
  • the B bands are centered at frequencies whose separation is monotonically non-decreasing.
  • the band separation is monotonically increasing in a log-like manner. Such a log-like manner is perceptually motivated.
  • they are on a psycho-acoustic scale, that is, the frequency bands are critically spaced, or follow a spacing related by a scale factor to critical spacing.
  • the banding of elements 109 and 115, and steps 209 and 217 is designed to simulate the frequency response at a particular location along the basilar membrane in the inner ear of a human.
  • the banding 109, 115, 209, 217 may include a set of linear filters whose bandwidth and spacing are constant on the Equivalent Rectangular Bandwidth (ERB) frequency scale, as defined by Moore, Glasberg and Baer (B. C. J. Moore, B. Glasberg, T. Baer, "A Model for the Prediction of Thresholds, Loudness, and Partial Loudness,” J. of the Audio Engineering Society (AES), Volume 45 Issue 4 pp. 224-240; April 1997).
  • ERP Equivalent Rectangular Bandwidth
  • the Bark frequency scale may be employed with reduced performance.
  • each of the single channels obtained for the mixed- down, e.g., beamformed input signals and for the reference input is reduced to a set of B spectral power (or other frequency domain amplitude metric), e.g., B such values on a psycho-acoustic scale.
  • B spectral power or other frequency domain amplitude metric
  • the B bands can be fairly equally spaced on a logarithmic frequency scale. All such log-like banding is called "perceptual banding" hereinin some embodiments, each band should have an effective bandwidth of around 0.5 to 2 ERB with one specific embodiment using a bandwidth of 0.7 ERB. In some embodiments, each band has an effective bandwidth of 0.25 to 1 Bark.
  • One specific embodiment uses a bandwidth of 0.5 Bark.
  • the inventors found it useful to keep the minimum band size to cover several frequency bins, as this avoids problems of temporal aliasing and circulant distortion in both time to frequency band— analysis— and frequency-to-time— synthesis— that can occur with transforms such as the short time Fourier transform. It is noted that certain transforms or subbanded filter banks such as the complex quadrature mirror filter, can avoid many of these issues.
  • the inventors found it advantageous that the characteristic shape and overlap of the banding used for power (or other frequency domain amplitude metric) representation and gain interpolation be relatively smooth.
  • the audio was high-pass filtered with a pass-band starting at around 100Hz. Below this, it was observed that the input, e.g., microphone signals are typically very noisy with a poor signal-to-noise ratio and it becomes increasingly difficult to achieve a perceptual spacing on account of the fixed length N transform. [00165] The bandwidth of a 1 ERB filter is given by
  • logarithmic banding for reasons related to the nature of hearing and perception, to achieve computational efficiency, and to improve the stability of statistical estimates across bands, the logarithmic banding is suggested and effective.
  • the logarithmic banding approach significantly reduces complexity and stabilizes the power estimation and associated processing that occur at higher frequencies.
  • the banding of elements 109, 115 and steps 209, 217 can be achieved with a soft overlap using banding filters, the set of banding filters also called an analysis filterbank.
  • the shape of each banding filter should be designed to minimize the time extent of the time domain filters associated with each band.
  • the banding operation of elements 109, 115 and steps 209, 217 can be represented by a B*N real- valued matrix taking the bin power (or other frequency domain amplitude metric) to the banded power (or other frequency domain amplitude metric). While not necessary, this matrix can be restricted to positive values as this avoids the problem of any negative band powers (or other frequency domain amplitude metric).
  • this matrix should be fairly sparse with bands only dependent on the bins around their center frequency.
  • An optimal filter shape for achieving the compact form in both the frequency and time domain would be a Gaussian.
  • An alternative with the same quadratic main lobe but a faster truncation to zero is a raised cosine. With each band extending to the center of the adjacent bands, the raised cosine also provides a unity gain when the bands are summed. Since the raised cosine becomes sharp for the smaller bands, it is advisable to also include an additional spreading kernel such as
  • this matrix is used to sum the powers (or other frequency domain amplitude metric) from the N bins into the B bands.
  • the transform of this matrix is used to interpolate the B suppression gains into a set of N gains to apply to the transform bins.
  • FIG. 5 depicts example shapes of the B bands in the frequency domain on both a linear and logarithmic scale. It can be seen that the B bands are approximately evenly spaced on the logarithmic scale with the lower bands becoming slightly wider. The term log-like is used for such behavior.
  • FIG. 5 Also shown in the FIG. 5 is the sum of example band filters. It can be seen that this has a unity gain across the spectrum with a high pass characteristic having a cut-off frequency around 100Hz.
  • the high frequency shelf and banding are not essential components of the embodiments presented herein, but are suggested features for use on typical microphone input signals for the case of the signal of interest being a voice input.
  • FIG. 6 shows time domain filter representations for several of the filter bands of example embodiments of banding elements 109, 115 and steps 209, 217.
  • an additional smoothing kernel [l 2 l]/ 4 is applied in the construction of the banding matrix coefficients. It can be seen that the filter extent is constrained to the center half of the time window around time zero. This property results by having the filter bands being wider than a single bin and, in this example, the additional smoothing kernel used in the determination of the banding matrix.
  • the property of constraining the filter extent to the center half of the time window has been found to reduce distortion due to circulant convolution when applying an arbitrary set of gains for the filter bank. This is of particular importance when using the same banding for both determining banded power (or other frequency domain amplitude metric) of signals, and for the operation shown in FIG. 3C of element 131, step 225 of interpolation used in applying the banded gains for the individual frequency bins.
  • step 225 of interpolation used in applying the banded gains for the individual frequency bins.
  • frequency domain amplitude metric representation is convenient in an implementation.
  • the analysis and interpolation banding may be different.
  • the inventors have found that constraining the filter extent to the center half of the time window is a particularly advantageous inherent in the banding matrix when used for interpolating the banded processing gains (element 131, step 225) to create binned gains to apply, when using the transform suggested above, or similar short term Fourier transform.
  • the banding of elements 109, 115 and steps 209, 217 serves several purposes: [00178] ⁇
  • ⁇ In some perceptual banding embodiments, psychoacoustic criteria are used for banding, and the resulting banding is related in some aligned or scaled way to the critical hearing bandwidth of a listener.
  • controlling the spectrum on a finer resolution than this has little merit, since the perceived activity in each band will be dominated by the strongest source in that band. The strongest source would also dominate the parameter estimation.
  • appropriate banding of the transform provides a degree of signal estimation and masking which matches inherent psychoacoustic models thus making use of masking in the suppression framework.
  • the spread of the bands on analysis and the gain constraint on output both work to avoid trying to suppress signal that is already masked. Smooth overlap of the bands provides further mechanism that effects a result similar to the computation of gains to achieve noise suppression that would take into account the a psychoacoustic masking effects of the listener.
  • a constraint can be applied to the banding design to ensure all the time domain filters related to the band filters have a compact form, with length ideally less than N. This design reduces distortion from circulant convolution when the band gains are applied in the transform domain.
  • some embodiments include scaling the power (or other metric of the amplitude) in each band to achieve some nominal absolute reference. This has been found useful for suppression in order to facilitate suppression of residual noise to a constant power across frequency value relative to the hearing threshold.
  • One suggested approach for normalization of the bands is to scale such that the 1kHz band has unity energy gain from the input, and the other bands are scaled such that a noise source having a relative spectrum matching the threshold of hearing would be white or constant power across the bands. In some sense, this is a pre-emphasis filter on the bands prior to analysis which causes a drop in sensitivity in the lower and higher bands.
  • This normalization is useful, since if the residual noise is controlled to be constant across the bands, this achieves a perceptually white noise when close to the hearing threshold. In this sense it provides a way of achieving sufficient but not excessive reduction of the signal by attenuating the bands to achieve a perceptually low or inaudible noise level, rather than just a numeric optimization in each band independent of the audibility of the noise.
  • T is the threshold of hearing in dB sound pressure level (SPL) which is
  • FIG. 7 shows the normalization gain for the banding to 30 bands as described above.
  • the 1kHz band is band 13 and thus has the OdB gain.
  • Y b is the banded instantaneous power of the mixed-down, e.g., beamformed signal
  • W b is the normalization gain from FIG. 7
  • w b n are the elements from the banding matrix shown in FIGS. 4 and 5.
  • the operation 217 of spectral banding element 115 forms X b , the banded instantaneous power of the combined reference signal, using the W b normalization gain and a banding matrix with elements w b n .
  • the prime notation can be generalized to any metric based on the frequency domain complex coefficients, in particular, their amplitude.
  • the 1-norm is used, i.e., the amplitude (also called envelope) of the spectral band is used, and the expression for the instantaneous mixed-down signal spectral amplitude becomes
  • useful metric is obtained by combining the weighted amplitudes across the bins used in a particular band, with exponent p, and then applying a further exponent of 1/q.
  • the goal of the method embodiments and system embodiments includes determining an estimate for the various components of the banded mixed-down audio signal that are included in the total power spectrum (or other amplitude metric spectrum) in that band. These are determined as power spectra (or other amplitude metric spectra). Determination of the components in a frequency band of the beamformed signal 3 ⁇ 4' is described below in more detail.
  • banded spatial feature estimator 105 determines whether banded spatial feature estimator 105 is a signal from the desired location and those not.
  • the beamformer 107 and beamforming step 207 may provide some degree of spatial selectivity. This may achieve some suppression of out-of -position signal power and some suppression of the noise and echo.
  • Suppression is carried out by applying a set of frequency dependent gains generally as real coefficients across the N frequency domain coefficients as suggested for embodiments presented herein.
  • the suppression gains are calculated in the banded domain from an analysis of signal features such as the power spectra (or other amplitude metric spectra).
  • i3 ⁇ 4' the total power spectrum (or other amplitude metric spectrum) of the banded mixed- down, e.g., beamformed signal power in band b.
  • FIGS. 8A and 8B show breakdowns of the various components in i3 ⁇ 4' , and the following is a brief description of the signal components in i3 ⁇ 4' with a discussion of assumptions associated with estimating the components in embodiments of the present invention.
  • N ⁇ is the power spectra (or other amplitude metric spectra) component which is reasonably constant or without short term flux, where flux, as is commonly understood by one skilled in the art, is a measure of how quickly the power spectrum (or other amplitude metric spectrum) changes over time.
  • Echo denoted 3 ⁇ 4 is the power spectra (or other amplitude metric spectra) component which has flux that is reasonably predictable given a short (0.25 - 0.5s) time window of the reference signal power spectra (or other amplitude metric spectra).
  • Out-of-position power denoted Power Ou t OfBeam > a ⁇ so called out-of-beam power and out-of-location power.
  • This is defined to be the power or power spectra (or other amplitude metric spectra) component with flux that does not have an appropriate phase or amplitude mapping on the input microphone signals to be potentially incident from the desired location.
  • FIG. 8A and FIG. 8B show two decompositions of the signal power (or other
  • FIG. 8A shows a separation of the echo power and noise power from power spectrum estimate of the mixed-down, e.g., beamformed signal to residual signal power, and further a separation into the desired in-position signal as a fraction of the residual signal power.
  • FIG. 8B shows a spatial of the total power in a band b into the total in-position power, and the total out-of-position power, and a separation of the total in-position power to an estimate of the desired signal power without an in-position echo power component and an in-position noise power component from the in-position power.
  • Embodiments of the present invention use the available information used to create some bounds for the estimate of the power in the desired signal, and create a set of band gains accordingly that can be used to affect simultaneous combined suppression.
  • the desired signal power is 1) bounded from above by the residual power, i.e., the total power P b less the noise power N b and less the echo power E b , and 2) bounded from above by the portion of the total power P b that is estimated to be in-position, i.e., the part that is not out-of-position power Power'o ut o fleam- Estimating signal spectrum P b ' (element 121, step 211)
  • signal power (or other frequency domain amplitude metric) estimator 121 generates an estimate of the total signal power (or other metric of amplitude) in each band b.
  • Embodiments of the present invention include determining in element 121, step 211 the overall signal power spectra (or other amplitude metric spectra) and noise power spectra (or other amplitude metric spectra).
  • the mixed-down e.g., beamformed instantaneous signal power
  • the downmixing e.g., beamforming 207 is a linear and time invariant process for the duration of interest, the mapping of the statistic of the noise and echo from the inputs X n to the output of the downmixer, e.g., beamformer
  • the initial beamformer is a linear and time invariant process over the time of observation used for the estimation of statistics, e.g., the power spectra, and thus the nature of the estimates relative to the underlying signal conditions prior to the beamforming are not changing due to rapid adaption of the beamformer with the signal conditions.
  • the variance of such an estimate depends on the length of time over which the signal is observed. For longer transform blocks, e.g., N>512 at 16kHz, the immediate band power (or other frequency domain amplitude metric) suffices.
  • one embodiment determines the power estimate i3 ⁇ 4' using a first order filter to smooth the signal power (or other frequency domain amplitude metric) estimate.
  • i3 ⁇ 4' the total power spectrum estimate in band b carried out in estimator 121 , step 211 is
  • Pbp ⁇ y is a previously, e.g., the most recently determined signal power (or other frequency domain amplitude metric) estimate
  • ccp 3 ⁇ 4 is a time signal estimate time constant
  • F min is an offset.
  • Pbp ⁇ y is a previously, e.g., the most recently determined signal power (or other frequency domain amplitude metric) estimate
  • ccp 3 ⁇ 4 is a time signal estimate time constant
  • F min is an offset.
  • Alternate embodiments use a different smoothing method, and may not include the offset.
  • a suitable range for the signal estimate time constant ctp 3 ⁇ 4 was found to be between 20 to 200 ms.
  • a narrower range of 40 to 120ms is used in some embodiments.
  • the offset F mm is added to avoid a zero level power spectrum (or other amplitude metric spectrum) estimate.
  • F min can be measured, or can be selected based on a priori knowledge.
  • F min for example, can be related to the threshold
  • the instantaneous power (or other frequency domain amplitude metric) 3 ⁇ 4 is a sufficiently accurate estimate of the signal power (or other frequency domain amplitude metric) spectrum i3 ⁇ 4' , such that element 121 is not used, but 3 ⁇ 4' is used for i3 ⁇ 4' .
  • the banding filters and the frequency bands are chosen according to criteria based on psycho-acoustics, e.g., with the log-like banding as described above. Therefore, in the formulae presented herein in which i3 ⁇ 4 is used, some embodiments use 3 ⁇ 4' instead.
  • Method 200 includes step 221 of performing prediction of the echo using adaptively determined echo filter coefficients (see echo spectral prediction filter 117), performing noise spectral estimation using the predicted echo spectral content and the total signal power (see noise estimator 123), updating the voice-activity echo detector (VAD) using the signal spectral content, noise spectral content, and echo spectral content (see element 125), and adapting the echo filter coefficients based on the VAD output and the signal spectral content, noise spectral content, and echo spectral content (see adaptive filter updater 127 that updates the coefficients of filter 117).
  • Instantaneous echo prediction of element 117 (part of step 221)
  • the echoes are created at the microphones due to the acoustic reproduction of signals related to the one or more reference signals.
  • the potential source of echoes are typically rendered, e.g., via a set of one or more loudspeakers.
  • a summer 111 is used to determine a direct sum of the Q rendered reference signals to generate a total reference to be used for echo spectral content prediction for suppression.
  • such a sum or grouped echo reference may be obtained by a single non-directional microphone having a much greater level of echo and lower level of the desired signal compared to the signals of input microphones.
  • the signals are available in pre-rendering form.
  • the digital signals that are converted to analog then rendered to a set of one or more loudspeakers may be available.
  • the analog speaker signals may be available.
  • the electronic signals, analog or digital are used, and directly summed by a summer 111 , in the digital or analog domain to provide M-sample frames of a single real-valued reference signal. The inventors have found that using the signals pre-rendering provides advantages.
  • Step 213 of method 200 includes the accepting (and summing) of the Q reference signals.
  • Step 215 includes transforming the total reference into frequency bins, e.g., using a time-to-frequency transformer 113 or a processor running transform method instructions.
  • Step 217 includes banding to form B spectral bands of the transformed reference, e.g., using a spectral bander 115 to generate the transform instantaneous power or other metric denoted X b ' . This is used to predict the echo spectral content using an adaptive filter.
  • the adaptive filter includes determining the instantaneous echo power spectrum (or other amplitude metric spectrum), denoted T b for band b by using an L tap adaptive filter described by
  • n ⁇ F b X b ' ,
  • filter coefficients are determined by an adaptive filter coefficient updater 127.
  • the filter coefficients require initialization, and in one embodiment, the coefficients are initialized to 0, and in another, they are initialized to an a priori estimate of the expected echo path.
  • One option is to initialize the coefficients to produce an initial echo power estimate that has a relatively high value - larger than any expected echo path which facilitates an aggressive starting position for echo and avoids the problem of an underestimated echo triggering the VAD and preventing adaption.
  • Adaptively updating the L filter coefficients uses the signal power (or other frequency domain amplitude metric) spectrum estimate P b from the current time frame and the noise power (or other frequency domain amplitude metric) spectrum estimate N from the current time frame.
  • Y b is a reasonably good estimate of P b , so is used for determining the L filter coefficients rather than P b (which in any case is determined from
  • One embodiment includes time smoothing of the instantaneous echo from echo
  • E b' a E,b T b + i 1 - E,b ) E bPr e V for T b ⁇ E bPr ev
  • 3 ⁇ 4p r ev is the previously determined echo spectral estimate, e.g., in the most recently, or other previously determined estimate
  • ⁇ 3 ⁇ 4 b is a first order smoothing time constant
  • the time constant in one embodiment is not frequency-band-dependent, and in other embodiments is frequency-band dependent. Any value between 0 and 200ms could work. A suggestion for such time constants ranges from 0 to 200ms and in one embodiment the inventors used values of 15 to 200 ms as a frequency-dependent time constant embodiments, whilst in another a non-frequency-dependent value of 30ms was used.
  • N b The noise power spectrum (or other amplitude metric spectrum) denoted N b is
  • a simple noise estimation algorithm can provide appropriate performance.
  • One example of such an algorithm is the minimum statistic. See R. Martin, "Spectral Subtraction Based on Minimum Statistics," in Proc. Euro. Signal Processing Conf. (EUSIPCO), 1994, pp. 1182-1185. Using the minimum statistic (a minimum follower) is appropriate, e.g., when the signal of interest has high flux and drops to zero power in any band of interest reasonably often, as is the case with voice.
  • one embodiment of the invention includes echo-gated noise estimation: updating the noise estimate N , and stopping the update of the noise estimate when the predicted echo level is significant compared with the previous noise estimate. That is, that noise estimator 123 provides an estimate which is gated when the predicted echo spectral content is significant compared to the previously estimated noise spectral content.
  • a simple minimum follower based on a historical window can be improved. The estimate from such a simple minimum follower can jump suddenly as extreme values of the power enter and exit the historical window.
  • Some embodiments of the present invention use a "leaky” minimum follower with a tracking rate defined by at least one minimum follower leak rate parameter.
  • the "leaky” minimum follower has exponential tracking defined by one minimum follower rate parameter.
  • the noise spectral estimate is determined, e.g., by element 123, and in step 221 by a minimum follower method with exponential growth.
  • the minimum follower is gated by the presence of echo comparable to or greater than the previous noise estimate.
  • b is a parameter that specifies the rate over time at which the minimum follower can increase to track any increase in the noise.
  • the parameter b is best expressed in terms of the rate over time at which
  • That rate can be expressed in dB/sec, which then provides a mechanism for determining the value of b .
  • the range is 1 to 30dB/sec. In one embodiment, a value of 20dB/sec is used.
  • the one or more leak rate parameters of the minimum follower are controlled by the probability of voice being present as determined by voice activity detecting (VAD). If the probability of voice suggests there is a higher probability of voice being present, the leakage is a bit slower, and if there is probability there is not voice, one leaks faster. In one embodiment, a rate of lOdB/sec is used when there is voice detected, whilst a value of 20dB/sec is used otherwise.
  • VAD voice activity detecting
  • VADs may be used, and as described in more detail further in this description, one aspect of the invention is the inclusion of a plurality of VADs, each controlled by a small set of tuning parameters that separately control sensitivity and selectivity, including spatial selectivity, such parameters tuned according to the suppression elements in which the VAD is used in.
  • VAD Voice activity detector
  • VAD element 125 determines an overall signal activity level denoted S as
  • the measure S is a measure indicative of the number of bands that have a signal (indicated by Y b ' ) that exceeds the present estimate of noise and echo by pre-defined amounts, indicated by ⁇ ⁇ , ⁇ ⁇ > 1. Since the noise estimate is an estimate of the stationary or constant noise power (or other frequency domain amplitude metric) in each band, rather than being a true "voice" activity measure, the measure S is a measure of transient or short time signal flux above the expected noise and echo.
  • the VAD derived in the echo update voice-activity detector 125 and filter updater 127 serves the specific purpose of controlling the adaptation of the echo prediction.
  • a VAD or detector with this purpose is often referred to as a double talk detector.
  • the values of ⁇ ⁇ , ⁇ ⁇ are between 1 and 4. In a particular embodiment,
  • ⁇ ⁇ , ⁇ ⁇ are each 2.
  • Y' sens is set to be around expected microphone and system noise level, obtained by experiments on typical components. Alternatively, one can use the threshold of hearing to determine a value for Y sens .
  • Voice activity is detected, e.g., to determine whether or not to update the prediction filter coefficients in echo prediction filter coefficient adapter 127, by a threshold, denoted S thresh m tne value of S. In some embodiments a continuous variation in the rate of adaption may be effected with respect to S [00237]
  • the operation in the echo update voice activity detector 125 has been found to be a simple yet effective method for voice or local signal activity detection. Since ⁇ ⁇ > 1 and ⁇ ⁇ > 1 , each band must have some immediate signal content greater than the estimate of noise and echo. Typical values for ⁇ ⁇ , ⁇ ⁇ are around 2.
  • a signal to noise ratio of at least 3dB is required for a contribution to the signal level parameter S . If the current signal level is large relative to the noise and echo estimate, the summation term has a maximum of 1 for each band.
  • the sensitivity offset in the denominator of the expression for S prevents S and thus any derived activity detector, such as the VAD 125, from registering at low signal levels.
  • the summation over the B bands for S will thus represent the number of bands that have "significant" local signal. That is a signal not expected from the noise and echo estimates which are assumed to be reasonable once the system converges.
  • the suggested scaling related to band size and threshold of hearing creates an effective balancing of the VAD expression with each band having a similar sensitivity and perceptually weighted
  • VAD sensitivities to the various components of the overall signal strength
  • a location-specific VAD is used in some embodiments of gain calculator 129 and in gain calculating step 223.
  • Echo prediction filter coefficient adapter gated by an activity threshold
  • the echo filter coefficient updating of updater 127 is gated, with updating occurring when the expected echo is significant compared to the expected noise and current input power, as determined by the VAD 125 and indicated by a low value of local signal activity S.
  • the adaptive filter coefficients are updated as:
  • F b F b + s ⁇ s thresh '
  • ⁇ ⁇ is a tuning parameter tuned to ensure stability between the noise and echo estimate.
  • a typical value for ⁇ ⁇ is 1.4 (+3dB).
  • a range of values 1 to 4 can be used, ⁇ is a tuning parameter that affects the rate of convergence and stability of the echo estimate.
  • ⁇ 0.1 independent of the frame size M.
  • X s ' ens is set to avoid unstable adaptation for small reference signals.
  • X s ' ens is related to the threshold of hearing.
  • X s ' ens is a pre-selected number of dB lower than the reference signal, so is set relative to the expected power (or other frequency domain amplitude metric) of the reference signal, e.g., 30 to 60 dB below the expected power (or other frequency domain amplitude metric) of X3 ⁇ 4 in the reference signal.
  • S thresh it is 30dB below the expected power (or other frequency domain amplitude metric) in the reference signal.
  • the choice of value for S thresh depends on the number of bands. S thresh is between 1 and B, and for one embodiment having 24 bands to 8kHz, a suitable range was found to be between 2 and 8, with a particular embodiment using a value of 4.
  • a lower threshold could prevent the adaptive filter from correctly tracking changes in the echo path, as the echo estimate may be lower than the incoming echo and adaption would be prevented.
  • a higher threshold would allow faster initial convergence, however since a significant local signal would be required to cause a detection from the echo prediction control VAD 125, the filter updates will be corrupted during double talk.
  • a band-dependent weighting factor can be introduced into the echo update voice- activity detector 125 such that the individual band contributions based on the instantaneous signal to noise ratio are weighted across frequency for their contribution to the detection of signal activity.
  • perceptual-based e.g., log-like banding
  • the inventors have found it acceptable to have a uniform weighting.
  • a band-dependent weighting function can be introduced.
  • the updating is a very low complexity but effective approach for controlling the adaption and predicting the echo level.
  • the approach was also found to be fairly effective at avoiding bias in the noise and echo estimates caused by the potentially ambiguous joint estimation.
  • the proposed approach effectively deals with the interaction between the noise and the echo estimates and has been found to be robust and effective in a wide range of applications.
  • Even though the approach is somewhat unconventional, in that the noise estimation method and echo prediction methods may not be the most accepted and established methods known, the approach was found to work well, and allows simple but robust techniques to be used in a systematic way to effectively reduce and control any error or bias.
  • the invention is not limited to the particular noise estimation method used or to the particular echo prediction method used.
  • a solution to this problem is to force adaption initially or repeatedly when some reference signal commences, or initialize the echo filter to be the expected of upper bound of the expected echo path.
  • the echo power spectrum (or other amplitude metric spectrum) is estimated, and this estimate has a resolution in time and frequency as set out by the transform and banding.
  • the echo reference need only be as accurate and have a similar resolution to this representation. This provides some flexibility in the mixing of the Q reference inputs as discussed above.
  • the inventors also found that there is also a toleration of gain variation of around 3-6dB due to the suppression rule and suggested values of the echo estimate scaling used in the VAD and suppression formulae.
  • Some embodiments of the invention do not include echo suppression, only
  • the elements involved in generating the echo estimate might not be present, including the reference inputs, elements 111, 113, 115, filter 117, echo update VAD 125 and element 127. Furthermore, with respect to FIG. 2, steps 213, 215, 217, and 221 would not be needed, and step 223 would not involve echo suppression.
  • One aspect of embodiments of the invention is using the input signal data, e.g., input microphone data in the frequency or transform domain from input transformers 103 and transforming step 203 to form estimates of the spatial properties of the sound in each band. This is sometimes referred to as inferring the source direction or location.
  • the presence of near-field objects means that the spatial location of an object can only be expressed in terms of the expected signal properties at the array of sound arriving from that desired or other source.
  • the source position location is not determined, but rather characteristics of the incident audio in terms of a set of signal statistics and properties are determined as a measure of the probability of a source of sound being or not being at a particular location.
  • Embodiments of the present invention include estimating or determining banded spatial features, carried out in the system 100 by banded spatial feature estimator 105, and in method 200 by step 205. Some embodiments of the present invention use an indicator of the probability of the energy in a particular band b having originated from a spatial region of interest. If, for example, there is a high probability in several bands, it is reasonable to infer that is it from a spatial region of interest.
  • Embodiments of the present invention use spatial information in the form of one or more measures determined from one or more spatial features in a band b that are monotonic with the probability that the particular band b has such energy incident from a spatial region of interest. Such quantities are called spatial probability indicators.
  • position is used to refer to an expected relationship between the signals at the microphone array. This is best viewed as a "position" in the array manifold that represents all of the possible relationships that may occur between signals from the microphone array given different incident discrete sounds. Whilst there will be a definitive mapping between the "position" of a source in the array manifold, and its physical position, it is noted that the technique and invention herein do not rely in any way on this mapping being known, deterministic or even constant over time.
  • the P sets of N complex values after the microphone input transforms are routed to a processing element for banded positional estimation.
  • microphones in each transform bin can be used to infer some positional information about the dominant source in that frequency bin for the given processing instant.
  • it is possible to resolve the direction or position of at most P - 1 sources assuming that we know the number of sources. See, for example, Wax, M. and I. Ziskind, On unique localization of multiple sources by passive sensor arrays. IEEE Trans. Acoustics, Speech, and Signal Processing, vol. 37, no. 7, pp. 996-1000, 1989.
  • Such classical statistical methods are concerned with the numerical and statistical efficiency of the approach.
  • an approach is presented that provides a robust solution for the suppressive control of audio signals to achieve good subjective results rather than to optimize simpler objective criteria.
  • an estimate is made of a measure monotonic with the probability that energy in a given band at that point time could reasonably have arrived from the desired location, which is represented by a target position in the array manifold.
  • the target position in the array manifold may be based on a priori information and estimates, or it may take advantage of previous online estimates and tracking (or a combination of both).
  • the result of the spatial inference is to create an estimate for a measure of probability, e.g., as an estimated fraction or as an appropriate gain that relates to the estimated amount of signal from the desired location, in that band at that point in time.
  • one or more spatial probability indicators are determined in step 205 by banded spatial feature estimator 105, and used for suppression. These one or more spatial probability indicators are one or more measures in a band b that are monotonic with the probability that the particular band b has such energy in a region of interest.
  • the spatial probability indicators are functions of one or more weighted banded covariance matrices of the inputs.
  • the w b n provide an indication of how each bin is weighted for contribution to the bands. This creates an estimate of the instantaneous array covariance matrix at a given time and frequency instant. In general, with multi-bin banding, each band contains a contribution from several bins, with the higher frequency bands having more bins. This use of banded covariance has been found to provide a stable estimate of the covariance, such covariance being weighted to the signal content having the most energy.
  • the one or more covariance matrices are smoothed over time.
  • the banding matrix includes time dependent weighting for a weighted moving average, denoted as W3 ⁇ 4 / with elements w b n , where / represents the time frame, so that, over L time frames, [00260] [ ⁇ 1, ⁇ - ⁇ , ⁇
  • the smoothing is defined by a frequency dependent time constant a b :
  • R' b R a b R b + (l- )Rfe pr ev .
  • R' bp is a previously determined covariance matrix.
  • each band covariance matrix R'3 ⁇ 4 is a 2x2 Hermetian positive definite matrix
  • Rb' ll Rb'l2 ' where the overbar is used to indicate the complex conjugate.
  • the spatial features include a "ratio" spatial feature, a "phase” spatial feature, and a “coherence” spatial feature. These features are used to determine an out-of-location signal probability indicator, expressed as a suppression gain, and determined using two or more of the spatial features, and a spatially-selective estimate of noise spectral content determined using two or more of the spatial features. In some the embodiments described herein, the three spatial features ratio, phase, and coherence are used, and how to modify these embodiments to include only two of the spatial features would be
  • ratio a quantity that is monotonic with the ratio of the
  • Ratio b ' 101og 10 ⁇ 1 1 + ⁇ where ⁇ is a small offset added to avoid singularities, ⁇ can be thought of as the smallest expected value for R ⁇ j .
  • it is the determined, or estimated (a priori) value of the noise power (or other frequency domain amplitude metric) in band b for the microphone and related electronics. That is, the minimum sensitivity of any preprocessing used.
  • Rb2 ⁇ R b ⁇ 2 Rb2 ⁇ R b ⁇ 2 .
  • related measures of coherence could be used such as
  • the coherence feature is [00272]
  • FIG. 9A, 9B and 9C show the probability density functions over time of the spatial features Ratio Phase and Coherence ' ⁇ , respectively, for diffuse noise, shown solid, and a desired signal, in this case voice, shown by dotted lines, as calculated for two inputs captured by a two-microphone headset with a microphone spacing of around 50mm across 32 frequency bands.
  • the incoming signals were sampled at a sampling rate of 8kHz, and the 32 bands are on an approximate perceptual scale with center frequencies from 66Hz to 3.8kHz.
  • the expected ranges are -10 to +10dB for Ratio' h , -180° to 180° for
  • Plots such as shown in FIGS. 9A, 9B and 9C are useful for determining the design of the probability indicators, in that they represent the spread of feature values that would be expected for the desired and undesired signal content.
  • the noise field is diffuse and can be comprised of multiple sources arriving from different spatial locations.
  • the spatial features Ratio' ⁇ , Phase' b , and Coherence' ⁇ for the noise are characteristic of a diffuse or spatially random field.
  • the noise is assumed to be in the farfield whilst the desired signal— the voice— is in the nearfield, however this is not a requirement for the application of this method.
  • the microphones were matched such that the average ratio feature for the noise field is OdB, i.e., a ratio of 1.
  • Noise signals arrive at the two microphones with a relatively constant expected power. For low frequencies the microphone signals would be expected to be correlated due to the longer acoustic wavelength, and the ratio feature for noise is concentrated around OdB.
  • the acoustic signal at the microphones can become independent in a diffuse field, and thus a spread in the probability density function of the ratio feature for noise is observed with higher frequency bands.
  • the phase spatial feature for the diffuse noise field is centered around 0°.
  • the characteristic of the head and device design create a deviation from the theoretical spaced microphone diffuse field response.
  • the wavelength decreases relative to the microphone spacing and the ratio and phase features for the noise become more distributed as the microphones become independent in the diffuse field.
  • the desired source does not emanate from a single location in space; speech from a human mouth has a complex and even dynamic spatial characteristic.
  • some embodiments of the invention use suppression not focused on the spatial geometry, but rather the statistical spatial response of the array for the desired source, as reflected by statistics of spatial features. While a simple theoretical model might suggest that the ratio and phase features would assume a single value for the desired source in the absence of noise, as shown in FIGS. 9A-9B, the ratio and phase features exhibit different values and spread in each band. This a priori information is used to determine the appropriate parameters for the probability indicators that are derived from each single observation of the features. This mapping can vary for the specific spatial
  • the coherence spatial feature is not dependent on any spatial configuration. Instead, it is a measure of the coherence or the extent to which the signal at that moment is being created by a single dominant source. As can be seen from FIG. 9C, at higher frequencies where the bands cover more frequency bins from the transform, the coherence feature is effective at separating the desired signal (a single voice) from the diffuse and complex noise field.
  • the distributions of the noise and desired signal show a degree of separation. From such distributions, one aspect of embodiments of the invention is to use an observation of each of these features in a given band to infer a partial probability of the incident signal being in the desired spatial location. These partial probabilities are referred to as spatial probability indicators herein.
  • spatial probability indicators In some bands the distributions of a spatial feature for voice and noise are disjoint, and therefore it would be possible to say with a high degree of certainty if the signal in that band is from the desired spatial location. However, there is generally some amount of overlap and thus the potential for noise to appear to have the desired statistical properties at the array, or for the desired signal to present a relationship at the microphone array that would normally be considered noise.
  • One feature of some embodiments of the invention is that, based on the a priori
  • each spatial feature in each band can be used to create a probability indicator for the feature for the band b.
  • One embodiment of the invention combines two or more of the probability indicators to form a combined single probability indicator used to determine a suppression gain, which, along with the additional information from noise and echo estimation, leads to a stable and effective combined suppression system and method.
  • the combining works to reduce the over processing and "musical" artifacts that would otherwise occur if each feature was used directly to apply a control or suppression to the signal. That is, one feature of embodiments of the invention is to make an effective combined inference or suppressive gain decision using all information, rather than to achieve a maximum suppression or discrimination from each feature independently.
  • the probability indicators designed are functions that encompass the expected
  • some embodiments of the invention include simplifying the distributions to a set of parameters.
  • the a priori characterization of the feature distributions for spatial locations is used to infer a centroid, e.g. a mean and an associated width, e.g., variance of the spatial features for sound originating from the desired location. This offers advantages over using detailed a priori knowledge: simplicity, and avoiding the possibility that in practice an over reliance on detailed a priori information can create unexpected results and poor robustness.
  • the distributions of the expected spatial features for the desired location are modeled as a Gaussian distributions that present a robust way of capturing the region of interest for probability indicators derived from each spatial feature and band.
  • Three spatial probability indicators are related to these three spatial features, and are the ratio probability indicator, denoted RPI' ⁇ , the phase probability indicator, denoted ⁇ , and the coherence probability indicator, denoted CPI' ⁇ , with
  • the function f Rb (ARatio') is a smooth function. In one embodiment, the ratio
  • Width Ratio b is a width tuning parameter expressed in log units, e.g., dB.
  • Width Ratio £ is related to but does not need to be determined from the actual data such as in FIG. 9A. It is set to cover the expected variation of the spatial feature in normal and noisy conditions, but also needs only be as narrow as is required in the context of the overall system to achieve the desired suppression. It is noted that the features presented in the example embodiments herein are nonlinear functions of the covariance matrix, and as such, the expected distribution of the feature values in a mixture of desired signal and noise, is typically not linearly related to the features for each signal separately. The introduction of any noise may cause a bias and variance to the observation of the features for the desired signal. Recognizing this, the target and widths could be selected or tuned to match the expected distributions in likely noise conditions.
  • Width Rat i 0 is not necessarily obtained from data such as shown in FIG. 9A. In one embodiment, assuming a Gaussian shape, Width Ratio 3 ⁇ 4 is 1 to 5dB which may vary with the band frequency.
  • APhase ⁇ ' Phase - Phase target ⁇ m ⁇ Phase tai g et is determined from either prior estimates or experiments on the equipment used, e.g., headsets, obtained, e.g., from data such as shown in FIG. 9B.
  • the function fp b (APhase') is a smooth function.
  • APIs' is a smooth function.
  • Width phase b i is related to but does not need to be determined from the actual data such as in FIG. 9B. It is set to cover the expected variation of the spatial feature in normal and noisy conditions, but also needs only be as narrow as is required in the context of the overall system to achieve the desired suppression. It typically needs to be tuned in the context of overall system performance.
  • the variance of the desired signal spatial features from sample data is a useful indication for the widths.
  • the spatial features are typically more stable, and therefore the widths could be narrow. Note however that too narrow a width may be overly aggressive, offering more suppressive ability than may be required at the expense of reduced voice or desired signal quality.
  • Matching the stability and selectivity of the spatial probability indicators is a process of tuning, guided by plots such as those of FIGS. 9 A and 9B, to achieve the desired performance.
  • the targets and widths for the ratio and phase features can be derived directly from data such as shown in FIG. 9A and 9B.
  • the targets may be obtained as the mean of the desired signal feature in each band, and the widths obtained from a scaling function of the variance of the same feature.
  • the targets and widths may be initially derived from data such as shown in FIGS. 9A and 9B and then adjusted as required to achieve a balance of noise reduction and performance.
  • CFactor ⁇ is a tuning parameter that may be a constant value in the range of 0.1 to 10; in one embodiment value of 0.25 was found to be effective.
  • CFactor j may dependent on frequency b, and typically have a lower value with increasing frequency b, e.g., with a range of up to 10 at low frequencies and decreasing to value 0 at the upper bands. In one embodiment, a value of about 5 is used for the lowest b, and a value of about 0.25 for the highest b.
  • Each of the probability indicators has a value between 0 and 1.
  • WidthUpR at(0 3 ⁇ 4 and WidthLow R at i 0 3 ⁇ 4.
  • WidthUp 3 ⁇ 4aie b and WidthDown 3 ⁇ 4aie b In one embodiment, f Phase b ' - Phase - target b
  • mapping from spatial feature to spatial probability indicators provide several useful examples. It should be evident that a set of curves could be created from any piecewise continuous function. By convention, the inventors chose that there should be at least some point or part of the spatial feature domain where the probability indicator is unity, with the function non-increasing as the distance from this point increases in either direction. For stable noise suppression and improved voice quality, the functions should be continuous and relatively smooth in value and also in the first and higher derivatives. Suggested extensions to the functions presented above include a "flat top" windowed region of the particular spatial feature, and other banded functions such as a raised cosine.
  • one embodiments includes determining pairwise spatial features and probability indicators for some or all pairs of signals. For example, for three microphones, there are three possible pairwise combinations. Therefore, for the case of determining the ratio, phase, and coherence spatial features, up to nine pairwise spatial features can be obtained, and probability indicators determined for each, and a combined spatial probability indicator determined for the configuration by combining two or more, up to nine spatial probability indicators.
  • the signal-of-interest position can be inferred along with such spatial features as a measure of uncertainty based on the coherence of the position across the transform bins associated with the given frequency band. If an assumption is made that the spectra of the sources creating the acoustic field are fairly constant across the transform bins in the frequency band, then each bin can be considered as a separate observation of the same underlying spatial distribution process.
  • One feature of embodiments of the invention is the use of statistical spatial
  • the gain calculator 129 uses the predicted echo spectral content, the instantaneous banded mixed-down signal power, together with the location probability indicators to implement one or more spatially-selective voice activity detectors, and to determine sets of B suppression probability indicators, in the form of suppression gains for forming a set of B gains for simultaneous noise, echo, and out-of-location signal suppression.
  • the suppression gain for noise (and echo) suppression uses a spatially-selective noise spectral content estimate determined using the location probability indicators. Beam gain and out-of-beam gain
  • One set of B gains is the beam gain, a probability indicator used to determine a
  • suppression probability indicator related to the probability of a signal coming from a source in the desired location or "in beam.” Similarly, related to this is a probability or gain for out- of-location signals, expressed in one embodiment as an out-of-beam gain.
  • the spatial probability indicators are used to determine what is referred to as the beam gain, a statistical quantity denoted BeamGain that can be used to estimate the in-beam and out-of-beam power from the total power, and further, can be used to determine the out-of-beam suppression gain.
  • the beam gain is the product of spatial probability indicators.
  • the probability indicators are scaled such that the beam gain has a maximum value of 1.
  • the beam gain is the product of at least two of the three spatial probability indicators.
  • the beam gain is the product of all three spatial probability indicators and has a maximum value of 1. Assuming each spatial probability indicator has a maximum value of 1 , in one embodiment, the beam gain has a pre-defined minimum value denoted BeamGain min This minimum serves to avoid the rapid fall of the beam gain to very low values where the variation in the gain value represents largely noise and small variations away from the signal of interest.
  • BeamGain BeamGain m ⁇ n + (1- BeamGain m ⁇ n )RPI b'PPI b'CPI 3 ⁇ 4.
  • Embodiments of the present invention use BeamGain min of 0.01 to 0.3 (-40dB to - lOdB).
  • One embodiment uses a BeamGain min of 0.1.
  • the beam gain is monotonic with the product of two or more of the spatial probability indicators.
  • one embodiment uses pairwise-determined spatial probability indicators, and in such an embodiment, the beam gain is monotonic with the product of the pairwise-determined spatial probability indicators.
  • the approach presented herein provides a simple method of combining the individual spatial feature probability indicators as a product and applying a lower threshold. The invention, however is not limited to such a combining.
  • Alternative embodiments of combining include one or more of using the maximum, minimum, median, average (on log or linear domain) or, with larger numbers of features with more than two inputs, an approach such as a voting scheme is possible.
  • the beam gain is used to determine the overall suppression gain as described herein below.
  • the beam gain is also used in some embodiments to estimate the in-beam power (or other frequency domain amplitude metric), that is, the power (or other frequency domain amplitude metric) in a given band b likely to be from the location of interest, and the out-of- beam power— the power (or other frequency domain amplitude metric) in a given band b likely to not be from the location of interest.
  • location or the general idea of a spatial position and mapping to a particular location on an array manifold, might be at a different angle of arrival, or might be nearfield vs. farfield, and so forth.
  • Y b the total banded power (or other frequency domain amplitude metric) from the mixed-down inputs, i.e., after beamforming.
  • the in-beam and out-of beam powers are:
  • Powers jnBeam BeamGain'fr 2 Y
  • Power' b 0ut0fBeam (l-BeamGain' b 2 ) Y .
  • Power' b n B ea m an ⁇ Power' b,OutOfleam are statistical measures used for suppression. Out of beam power and a spatially-selective noise estimate
  • Embodiments of the present invention include determining an estimate of noise
  • noise is usually assumed to be stationary, whereas voice is assumed to have a high flux.
  • a spectrally monotonous voice signal might therefore be interpreted as noise, and should the suppression be based on such a noise estimate, there is a possibility that the voice will eventually be suppressed. It is desired to be less-sensitive to noise-like sounds that come from a location of interest.
  • a feature of some embodiments of the invention is use of the spatial probability indicators to improve the estimate noise power (or other frequency domain amplitude metric) spectral estimate for use to determine suppression gains taking location into account in order to reduce the sensitivity of suppression to noise-like sounds that come from a location of interest.
  • the noise suppression gain is based on a spatially-selective estimate of noise spectral content.
  • Another feature of some embodiments is the use of the spatial probability indicators to carry out spatially sensitive voice activity detection, which is used in carrying out suppression gains taking location into account.
  • interpreting voice as noise is not necessarily a disadvantage, e.g., for echo prediction control.
  • the noise estimate N y determined for voice activity detection and for updating the echo prediction filter doe not take location into account (except for any location sensitivity inherent in the initial beamforming).
  • FIG. 10 shows a simplified block diagram of an embodiment of the gain calculator 129 and includes a spatially- selective noise power (or other frequency domain amplitude metric) spectrum calculator 1005 that operates on an estimate of the out-of-beam power, denoted Power' OutOfBeam* generated by an out-of-beam power spectrum calculator 1003.
  • a spatially- selective noise power or other frequency domain amplitude metric
  • FIG. 11 shows a flowchart of gain calculation step 223, and post-processing step 225 in embodiments that include post-processing, together with the optional step 226 of calculating and incorporating an additional echo gain.
  • the out-of-beam power spectrum calculator 1003 determines the beam gain
  • BeamGain' b from the spatial probability indicators.
  • BeamGain BeamGain ' min + (1- BeamGain ⁇ RPI ⁇ -PPI ⁇ -CPIi,. [00332] Each of element 1003 and step 1105 determines an estimate of the out-of-beam
  • the out-of-beam banded spectral estimate and the out-of-beam banded spectral estimate are determined using the signal power (or other frequency domain amplitude metric) spectrum, P b , rather than Y b .
  • Y b is a good approximation of P b .
  • the inventors have found that if the spectral banding is sufficiently analytic, e.g., the banding is log-like and perceptually-based, then Y b is more or less equal to P b , and it is not necessary to use the smoothed power estimate P b .
  • One embodiment of the invention uses a leaky minimum follower, with a tracking rate determined by at least one or leak rate parameter.
  • the leak rate parameter need not be the same as for the non-spatially selective noise estimation used in the echo coefficient updating. [00339] Denote by N' 5 the spatially selective noise spectrum estimate 1006. In one embodiment,
  • the leak rate parameter ⁇ 3 ⁇ 4 is expressed in dB/s such that for a frame time denoted T, (l + b )VT is between 1.2 and 4 if the probability of voice is low, and 1 if the probability of voice is high.
  • the noise estimate is updated only if the previous noise estimate suggests the noise level is greater, e.g., greater than twice the current echo prediction. Otherwise the echo would bias the noise estimate.
  • Power b o OfBeam * s tne instantaneous quantity determined using Y while in another embodiment, the out-of-beam spectral estimate determined from P is used for calculating N' 5 .
  • the at least one leak rate parameter of the leaky minimum follower used to determine N' 5 are controlled by the probability of voice being present as determined by voice activity detecting. Noise suppression (possibly with echo suppression)
  • One aspect of the invention is simultaneously suppressing: 1) noise based on a
  • each of an element 1013 of gain calculator 129 and a step 1108 of step 223 calculates a probability indicator, expressed as a gain for the intermediate signal, e.g., the frequency bins 108 based on the spatially selective estimates of the noise power (or other frequency domain amplitude metric) spectrum, and further on the instantaneous banded input power Y b in a particular band.
  • a probability indicator is referred to as a gain, denoted Gain N .
  • this gain Gain N is not directly applied, but rather combined with additional gains, i.e., additional probability indicators in a gain combiner 1015 and in a combining gain step 1109 to achieve a single gain to apply to achieve a single suppressive action.
  • each of elements 1013 and step 1108 is shown in FIGS. 10 and 11, respectively, with echo suppression, and in some versions does not include echo suppression.
  • Y is the instantaneous banded power (or other frequency domain amplitude metric)
  • the banded spatially- selective (out of beam) noise estimate and ⁇ ⁇ ' is a scaling parameter, typically in the range of 1 to 4, to allow for error in the noise estimate and to offset the gain curve accordingly.
  • the parameter GainExp is a control of the aggressiveness or rate of transition of the suppression gain from suppression to transmission. This exponent generally takes a value in the range of 0.25 to 4 with a preferred value in one embodiment being 2.
  • Some embodiments of the invention include not only noise suppression, but
  • some embodiments of the invention include simultaneously suppressing: 1) noise based on a spatially selective noise estimate, 2) echoes, and 3) out-of-beam signals.
  • element 1013 includes echo
  • step 1108 include echo suppression.
  • the probability indicator for suppressing echoes is expressed as a gain denoted Gain ' ⁇ N+E .
  • the above noise suppression gain expression in the case of also including echo suppression, becomes
  • Gain b ⁇ N+E (“Gain ⁇ ) where 3 ⁇ 4' is again the instantaneous banded power, N3 ⁇ 4 s , E b are the banded spatially- selective noise and banded echo estimates, and ⁇ ⁇ ' , ⁇ ⁇ ' are scaling parameters in the range of
  • GainExp ⁇ in expression Gain 1 is a control of the aggressiveness or rate of transition of the suppression gain from suppression to transmission. This exponent would generally take a value in the range of 0.25 to 4 with a preferred value for one embodiment being 2 for all values of b. [00353] In the remainder of the section on suppression, echo suppression is included.
  • some embodiments of the invention do not include echo suppression, only simultaneous suppression of noise and out-of-location signals.
  • the elements involved in generating the echo estimate might not be present, including the reference inputs, elements 111, 113, 115, filter 117, echo update VAD 125 and element 127.
  • steps 213, 215, 217, and 221 would not be needed, and step 223 would not involve echo suppression.
  • Gain 1 for Gain N+E applicable to simultaneous noise and echo suppression
  • MMSE minimum mean squared error
  • the present invention is broader, and in embodiments of the present invention, value of the GainExp ⁇ larger than 0.5 is found to be preferable in creating a transition region between suppression and transmission that is more removed from the region of expected noise power activity and variation.
  • the gain expressions achieve a relatively flat, or even inverted gain relationship with input power in the region of expected noise power - and the inventors consider this an inventive step in the design of the gain functions that significantly reduces instability of the suppression during noise activity.
  • element 1013 and 1108 have the instantaneous banded input power (or other frequency domain amplitude metric) 3 ⁇ 4' in both the numerator and denominator. This works well when the banding is properly designed as described herein, with log-like or perceptually spaced frequency bands.
  • the denominator uses the estimated banded power spectrum (or other amplitude metric spectrum) i3 ⁇ 4' , so that the above expression for Gain ⁇ N+E changes to:
  • One feature of some embodiments of the invention is significantly reducing this problem.
  • FIG. 12 shows a probability density in the form of a scaled histogram of signal power in a given band for the case of noise (solid line) and desired (voice) signal (broken line) in isolation obtained from observing around 10s of each signal class for a single band of around 1kHz where the noise and voice level correspond to an average signal to noise level of around OdB.
  • the values are illustrative and not restrictive and it should be evident that this figure serves to capture the characteristics of the suppression gain calculation problem in order to demonstrate the desired properties and specific designs of some embodiments of such calculations.
  • the horizontal axes represent a scaled value of the instantaneous band power relative to the expected noise (and echo) power. This is effectively the ratio of input power to noise, which is related but slightly different to the more commonly used signal to noise ratio.
  • some lower limit must be placed on the noise and/or echo estimate such that the ratio of input signal power to noise remains bounded.
  • the value of this limit is not material, provided it is sufficiently small, since the probability indicators, expressed herein as gain functions, are asymptotically unity for large ratios of input power to expected noise.
  • the representation of gain vs. input power described herein is preferred to a more conventional representation in terms of gain vs. signal to noise ratio, as it better demonstrates the natural distribution of power in the different signal classes, and serves to highlight the design and benefits of using the gain expressions described herein.
  • expected noise and echo power' is used to refer to the sum of the expected noise power and expected echo power at that time. At any specific time in a band, there could be either echo or noise or both signals present in any proportion.
  • the noise signal shows a spread of observed instantaneous input signal powers centered around the noise estimate and having an approximate range of +10dB.
  • the desired signal in this case of voice, has a higher instantaneous power having a larger range and generally having an instantaneous power in the range of 5-20dB more than the noise when there is active voice.
  • the data was representative of an incident signal at the microphone where the ratio of the average voice signal and noise signal power was OdB.
  • OdB the ratio of the average voice signal and noise signal power
  • any suppression gain should attenuate the noise components by a constant, and transmit the speech with unity gain.
  • the distributions of the desired signal and noise are not disjoint.
  • the design criteria for suppression used work to ensure relatively stable gain across the most probable speech levels and the most probable noise levels in order to avoid artifacts being introduced.
  • this is a new non-obvious inventive way of posing, visualizing and achieving a superior performing outcome for the suppression system.
  • Many prior art approaches are concerned with minimizing the numerical error in each bin or band against the original reference, which can lead to unstable gains and musical artifacts common in other solutions.
  • One feature of embodiments of the invention is the specification of the suppression gains for each band in the form of properties of the gain functions.
  • the constant or smooth gains across both the voice and noise power distribution modes ensures processing and musical noise musical artifacts are significantly reduced.
  • the inventors have found also that the methods presented herein can reduce the reliance on accurate estimates for the noise and echo levels.
  • the minimum value selected, 0.1 is not meant to be limiting, and can be different in different embodiments.
  • the inventors suggest a range of from 0.001 to 0.3 (-60dB to -lOdB), and the minimum can be frequency dependent.
  • the second uses a softer additive minimum which achieves both a flatter gain around the expected noise/echo power and also a smoother transition and first derivative, e.g.,
  • the minimum value selected, 0.1 is not meant to be limiting, and can be different in different embodiments.
  • the inventors suggest a range of from 0.001 to 0.3 (-60dB to -lOdB), and the minimum can be frequency dependent.
  • the second value is sensibly 1 minus the first value.
  • GainExp' ⁇ is a parameter usable to control the aggressiveness of the transition from suppression to transmission and may take values ranging from 0.5 to 4 with a preferred value in one embodiment being 1.5.
  • the first two values, shown here as 0.1 and 0.01 are adjusted to achieve the required minimum gain value and transition period.
  • the minimum value shown, 0.1 is not meant to be limiting, and can be different in different embodiments.
  • the scalar 0.01 is set to achieve an attenuation of around 8dB with the input power at the expected noise and echo level. Again, different values can be used in different embodiments.
  • a fifth example presents a generalization of this using the well known logistic equation
  • FIG. 13 shows the distribution of FIG. 12, together with the gain expressions Gain 1, Gain 2, Gain 3, and Gain 4 described above as functions of the ratio of input power to noise. The gain functions are shown plotted on a log scale in dB.
  • features of this family of suppression gain functions include, assuming that for each frequency band, a first range of values of banded instantaneous amplitude metric values is expected for noise, and a second range of values of banded instantaneous amplitude metric values is expected for a desired input:
  • ⁇ A (relatively) constant gain for the first range of values, i.e., in the region of the noise power.
  • relatively constant is meant, e.g., less than 0.03 dB of variation in the range.
  • ⁇ A (relatively) constant gain for the second range of values, i.e., in the region of the desired signal, e.g., voice signal power.
  • relatively constant is meant, e.g., less than 0.1 dB per dB of input power in the second range.
  • ⁇ A (relatively) smooth transition from the first range to the second range, i.e., from the region of the noise power to the region of desired signal power.
  • ⁇ The progression towards a function whose derivative also is smooth, e.g., a
  • ⁇ A relatively smooth transition from the region of the noise power to the region of desired signal power.
  • ⁇ A continuous and bound first and desirably higher derivatives.
  • This approach substantially reduces the degree of expansion that may occur due to excessive gradient or discontinuities in the gain as a function of the incoming banded signal power.
  • instantaneous power of approximately -0.5 in units of dB gain vs. dB input power, where approximately means -0.3 to -0.7.
  • a slope of -0.5 is suggested and achieves a compression ratio of the dynamic range of the noise signal of 2:1.
  • a modified sigmoid function is used; the sigmoid function is modified by including an additional term to result in a desired negative gradient for input signal powers around the expected noise level.
  • a modified sigmoid function is used that includes a sigmoid function and an additional term to provide the negative gradient in the first region.
  • FIG. 14 shows the histograms of FIG. 12 together with the sigmoid gain curve of Gain 4 and the modified sigmoid-like gain curve of Gain 5, called the whitening gain on the drawing.
  • Each of the plots has the input power to noise ratio in dB as the horizontal axis.
  • FIG. 15 shows what happens to the probability density functions, shown as scaled histograms, for the expected power of the noise for a noise signal and for a voice signal after applying the sigmoid-like gain curve Gain 4 and the whitening gain Gain 5. As can be seen, each of these causes a significant increase in the separation of the voice and noise, with the noise level decreasing in power or shifting lower on the horizontal axis.
  • the first sigmoid gain, Gain 4 creates a spreading of the noise power. That is, the noise level fluctuates more in power than in the original noise signal. This effect may be worse for many prior art approaches to noise suppression that do not exhibit the smooth property of the sigmoid like functions through the main noise power distribution.
  • the voice levels are also slightly expanded.
  • the second modified sigmoid gain, Gain 5 has the property of compacting the noise power distribution. This makes the curve higher, since the central noise levels are now more probable. This means there are less fluctuations in the noise and a sort of smoothing or whitening which can lead to less intrusive noise.
  • these plots show scaled probability density functions, as histograms, for noise, and for a voice signal. The noise and voice probability density functions are scaled to have the same area.
  • both gain functions increase the signal to noise ratio by increasing the spread— reducing the noise levels.
  • the noise is less intrusive and partially whitened over time and frequency.
  • the undesirable signal power is the sum of the estimated (location-sensitive) noise power and predicted or estimated echo power. Combining the noise and echo together in this way provides a single probability indicator in the form of a suppressive gain that causes simultaneous attenuation of both undesirable noise and of undesirable echo.
  • f A ( ⁇ ) , f B ( ⁇ ) a pair of suppression gain functions, each having desired properties for suppression gains, e.g., as described above, including, for example being smooth.
  • each of f A ⁇ -) , g (-) has sigmoid function characteristics.
  • the suppression probability indicator for in-beam signals is the suppression probability indicator for in-beam signals
  • Gain ⁇ s is determined by a spatial suppression gain calculator 1011 in element 129 (FIG. 10) and by a calculating suppression gain step 1103 in step 223 as
  • the spatial suppression gain 1012 is combined with other suppression gains in gain combiner 1015 and combining step 1109 to form an overall probability indicator expressed as a suppression gain.
  • Gainb' 0.1 + 0.9Ga 3 ⁇ 4 5 ⁇ Gain ⁇ N+E .
  • f A achieves (relatively) modest suppression of both noise and echo, while f B —7- suppresses the echo more.
  • f A ( ⁇ ) suppresses only noise
  • /# ( ⁇ ) suppresses the echo.
  • Gain b ' RAW 0.1 + 0.9Gain b ' s ⁇ Gain b ' ⁇ N+E ,
  • this noise and echo suppression gain is combined with the spatial feature probability indicator or gain for form a raw combined gain.
  • the raw combined gain is post-processed by a post-processor 1025 and by post processing step 225 to ensure stability and other desired behavior.
  • the gain function specific to the echo suppression is the gain function specific to the echo suppression
  • gain calculator 129 includes a determined of the additional echo suppression gain and a combiner 1027 of the additional echo suppression gain with the post-processed gain to result in the overall B gains to apply. The inventors discovered that such an embodiment can provide a more specific and deeper attenuation of echo. Note that in embodiments that include post-processing, the echo probability indicator
  • the postprocessing 225 is not subject to the smoothing and continuity imposed by the postprocessing 225, such post-processing, e.g., being tailored for the desired signal and noise signal stability, and a suitable level of noise suppression without unwanted voice distortion.
  • the need to eliminate echo from the signal can override the constraint of instantaneous speech quality when echo is active.
  • the echo suppressive component (after post-processing in embodiments that include post-processing) can apply narrow and potentially deep suppressive action across frequency, which can leave an unpleasant residual signature of the echo on the remaining noise in the signal.
  • a solution to this problem is that of "comfort noise" and it should be well known to some-one skilled in the art, and apparent how this could be applied to reduce the presence of gaps in the spectrum caused by an echo suppressor after the gain post processing.
  • Some embodiments of the gain calculator 129 include a post-processor 1025 and some embodiments of method 200 include a post-processing step 225.
  • Each of the post processor and post-processing step 225 is to post process the combined raw gains of the bands to generate a post-processed gain for each band.
  • Such post-processing includes in different embodiments one or more of: ensuring minimum gain values; ensuring there are no or few isolated or outlier gains by carrying out median filtering of the combined gain; and ensuring smoothness by carrying out one or both of time smoothing and band-to-band smoothing.
  • Some embodiments include signal classification, e.g., using one or both: a spatially-selective voice activity detector 1021 implementing a step 1111 , and a wind activity detector 1023 implementing a step 1113 to generate a signal classification, such that the postprocessing 225 of post-processor 1025 is according to the signal classification.
  • FIG. 1021 An embodiment of a spatially-selective voice activity detector 1021 is described herein below, as is an embodiment of a wind activity detector (WAD) 1023.
  • the signal classification controlled post-processing aspect of the invention is not limited to the particular embodiments of a voice activity detector or of a wind activity detector described herein.
  • the raw combined gain Gain b may sometimes fall below a desired minimum point, that is, achieve more than a maximum desired suppression depth.
  • maximum suppression depth and minimum gain shall be uses interchangeably herein.
  • Not all the above-described embodiments for determining the gain include ensuring that the gain does not fall below such a minimum point.
  • the step of ensuring a minimum gain serves to stabilize the suppressive gain in noisy conditions by avoiding low gain values that can exhibit large relative variation with small errors in feature estimation or natural noise feature variations.
  • post-processor 1025 and post processing step 225 include, e.g., in step 1115, ensuring that the gain does not fall below a pre-defined minimum, so that there is a pre-defined maximum suppression depth.
  • post-processor 1025 and step 1115 rather than the raw gain having the same maximum suppression depth (minimum gain) for all bands, it may be desired that the minimum level be different for different frequency bands.
  • minimum level be different for different frequency bands.
  • Gain b RAW Gain b MIN + (l - Gain b MIN ) ⁇ Gain b s ⁇ Gain b ' ⁇ N+E .
  • the range of the maximum suppression depth or minimum gain may range from -80dB to -5dB and be frequency dependent.
  • the suppression depth was around -20dB at low frequencies below 200Hz, varying to be around -lOdB at 1kHz and relaxing to be only - 6dB at the upper voice frequencies around 4kHz.
  • the processing of post-processing step 225 and of postprocessor 1025 is controlled by a classification of the input signals, e.g., as being voice or not as determined by a VAD, and/or as being wind or not as determined by a WAD.
  • the minimum values of the gain for each band, Gain ⁇ MjN are dependent on a classification of the signal, e.g., whether the signal is determined to be voice by a VAD in embodiments that include a VAD, or to be wind by embodiments that include a WAD.
  • the VAD is spatially selective.
  • the amount of increase in the minimum is larger in the mid-frequency bands, e.g., bands between 500 Hz to 2kHz.
  • the increase in minimum gain values is controlled to increase in a gradual manner over time as voice is detected, and similarly, to decrease in a gradual manner over time as lack of voice is detected after voice has been detected.
  • the decrease in minimum gain values is controlled to decrease in a gradual manner over time as wind is detected, and similarly, to increase in a gradual manner over time as lack of wind is detected after wind has been detected.
  • a single time constant is used to control the increase or decrease (for voice) and the decrease or increase (for wind).
  • a first time constant is used to control the increase in minimum gain values as voice is detected or the decrease as wind is detected
  • a second time constant is used to control the decrease in minimum gain values as lack of voice is detected after voice was detected, or the increase in minimum gain values as lack of wind is detected after wind was detected.
  • Such statistical outliers might occur in other types of processing in which an input signal is transformed and banded.
  • Such other types of processing include perceptual domain- based leveling, perceptual domain-based dynamic range control, and perceptual domain- based dynamic equalization that takes into account the variation in the perception of audio depending on the reproduction level of the audio signal. See, for example, International
  • Perceptual-domain-based leveling, perceptual-domain-based dynamic range control, and perceptual-domain-based dynamic equalization processing each includes determining and adjusting the perceived loudness of an audio signal by applying a set of banded gains to a transformed and perceptually-banded metric of the amplitude of an input signal.
  • a psychoacoustic model is used to calculate a measure of the loudness of an audio signal in perceptual units.
  • perceptual domain loudness measure is referred to as specific loudness, and is a measure of perceptual loudness as a function of frequency and time.
  • true dynamic equalization is carried out in a perceptual domain to transform the perceived spectrum of the audio signal from a time- varying perceived spectrum to a substantially time-invariant perceived spectrum.
  • the gains determined for each band for leveling and/or dynamic equalization include statistical outliers, e.g., isolated values, and such outliers might cause artifacts such as musical noise.
  • the processing described herein may be applicable also to such other applications in which gains are applied to a signal indicative of transformed banded norms of the amplitude at a plurality of frequency bands.
  • the proposed post processing is also directly applicable to systems without the combination of features and suppression. For example, it provides an effective method for improving the performance of a single channel noise reduction system.
  • One embodiment of post-processing 225 and of post-processor 1025 includes, e.g. in step 1117, median filtering the raw gain over different frequency bands.
  • the median filter is characterized by 1) the number of gains to include to determine the median, and 2) the conditions used to extend the banded gains to allow calculation of the median at the edges of the spectrum.
  • One embodiment includes 3 -point band-to-band median filtering, with extrapolation of interior values for the edges.
  • the minimum gain or a zero value is used to extend the banded gains.
  • the band-to-band median filtering is controlled by the signal classification.
  • a VAD e.g., a spatially-selective VAD is included, and if the VAD determines there is no voice, 5-point band-to-band median filtering is carried out, with extending the minimum gain or a zero value at the edges to compute the median, and if the VAD determines there is voice present, 3-point band-to-band median filtering is carried out, extrapolating the edge values at the edges to calculate the median.
  • a WAD is included, and if the WAD determines there is no wind, 3-point band-to-band median filtering is carried out, with extrapolating the edge values applied at the edges, and if the WAD determines there is wind present, 5-point band-to-band median filtering is carried out, with selecting the minimum gain values applied at the edges.
  • post-processor 1025 and post-processing step 225 include smoothing 1119 across the bands to eliminate such potential jumps which can cause colored and unnatural output spectra.
  • One embodiment of smoothing 1119 uses a weighted moving average with a fixed kernel.
  • One example uses a binomial approximation of a Gaussian weighting kernel for the weighted moving average.
  • a 5-point binomial smoother has a kernel— [l 4 6 4 l] .
  • the factor 1/16 may be left out, with scaling carried out in one point or another as needed.
  • a 3 -point binomial smoother has a kernel - ⁇ - [l 2 l] .
  • Many other weighted moving average filters are known, and any such filter can
  • the smoothing, e.g. of step 1119 can be defined by a real- valued square matrix of dimension B, the number of frequency bands.
  • each of the gain applications of element 131 and the step 227 incorporates band-to-band smoothing.
  • the band-to-band median filtering is controlled by the signal classification.
  • a VAD e.g., a spatially-selective VAD is included, and if the VAD determines there is voice, the degree of smoothing is increased when noise is detected.
  • 5-point band-to-band weighted average smoothing is carried out in the case the VAD indicates noise is detected, else, when the VAD determines there is no voice, no smoothing is carried out.
  • time smoothing of the gains also is included. In some embodiment
  • the gain of each the B bands is smoothed by a first order smoothing filter:
  • Gain btSmoothed a b Gain b + ⁇ l - a b )Gain bf Smoothedpi ev
  • Gain b is the current time-frame gain
  • Gain ⁇ Smoothedpr ev is Gain b Smoothed from the previous M-sample frame.
  • a b is a time constant which may be frequency band dependent and is typically in the range of 20 to 500ms. In one embodiment a value of 50ms was used.
  • first order time smoothing of the gains according to a set of first order time constants is included.
  • the amount of time smoothing is controlled by the signal classification of the current frame.
  • the signal classification of the current frame is used to control the values set of first order time constants used to filter the gains over time in each band.
  • one embodiment stops time smoothing in the case voice is detected.
  • the parameters of post-processing are controlled by the immediate signal classifier (VAD, WAD) value that has low latency and is able to achieve a rapid transition of the post-processing from noise into voice (or other desired signal) mode.
  • VAD immediate signal classifier
  • WAD voice-based signal classifier
  • the speed with which more aggressive post-processing is reinstated after detection of voice, i.e., at the trail out, has been found to be less important, as it affects intelligibility of speech to a lesser extent.
  • VADs are known in the art.
  • so-called “optimal VADs” are known, and there has been much research on how to determine such an "optimal VAD” according to a VAD optimality criterion.
  • one aspect of the invention is the inclusion of a plurality of VADs, each controlled by a small set of tuning parameters that separately control sensitivity and selectivity, including spatial selectivity, such parameters tuned according to the suppression elements the VAD is used in.
  • Each of the plurality of the VADs is an instantiation of a universal VAD that
  • the universal VAD determines indications of voice activity from Y' b .
  • the universal VAD is controlled by a set of parameters and uses an estimate of noise spectral content, the banded frequency domain amplitude metric representation of the echo, and the banded spatial features.
  • the set of parameters includes whether the estimate of noise spectral content is spatially selective or not.
  • the type of indication of voice activity an instantiation determines controlled by a selection of the parameters.
  • another feature of embodiments of the invention is a method of determining a plurality of indications of voice activity from Y' j ,, the mixed-down banded instantaneous frequency domain amplitude metric, the indications using respective instantiations of a universal voice activity detection method.
  • the universal voice activity detection method is controlled by a set of parameters and uses an estimate of noise spectral content, the banded frequency domain amplitude metric representation of the echo, and the banded spatial features.
  • the set of parameters including whether the estimate of noise spectral content is spatially selective or not. Which indication of voice activity an instantiation determines controller by a selection of the parameters.
  • selectivity is important, that is, the VAD instantiation should have a high probability that what it is detecting is voice
  • sensitivity is important, that is, the VAD instantiation should have a low probability of missing voice activity, even at the cost of selectivity so that more false positives are tolerated.
  • the VAD 125 used to prevent updating of the echo prediction parameters—the prediction filter coefficients— is selected to have a high sensitivity, even at the cost of selectivity.
  • the inventors selected to tune a VAD to have a balance of selectivity and sensitivity as being overly sensitive would lead to fluctuation of levels in noise as speech was falsely detected, whilst being overly selective would lead to some loss of voice.
  • the measurement of output speech level requires a VAD that is highly selective, but not overly sensitive to ensure that only actual speech is used to set the level and gain control.
  • One embodiment of a general spatially selective VAD structure the universal VAD to calculate voice activity that can be tuned for various functions is
  • BeamGain' ⁇ BeamGain m ⁇ n + (1— BeamGain m ⁇ )RPI ⁇ b'CPI BeamGainExp is a parameter that for larger values increases the aggressiveness of the spatial selectivity of the VAD, and is 0 for a non-spatially selective VAD such as used for echo update VAD 125, ⁇ ', ⁇ Nb' $ denotes either the total noise power (or other frequency domain amplitude metric) estimate N as used in VAD 125, or the spatially selective noise estimate N3 ⁇ 4 s determined using the out-of-beam power (or other frequency domain amplitude metric), ⁇ ⁇ , ⁇ ⁇ > 1 are margins for noise end echo, respectively and Y s ' ens is a settable sensitivity offset.
  • ⁇ ⁇ , ⁇ ⁇ are between 1 and 4.
  • BeamGainExp is between 0.5 to 2.0 when spatial selectivity is desired, and is 1.5 for one embodiment of step 1111 and VAD 1021 used to control post-processing. [00461 ]
  • the above expression also controls the operation of the universal voice activity
  • a binary decision or classifier can be obtained by considering the test S > 3 ⁇ 4 riW 3 ⁇ 4 as indicating the presence of voice. It should also be apparent that the value S can be used as a continuous indicator of the instantaneous speech level.
  • an improved useful universal VAD for operations such as transmission control or controlling the post processing could be obtained using a suitable "hang over" or period of continued indication of voice after a detected event. Such a hang over period may vary from 0 to 500ms, and in one embodiment a value of 200ms was used. During the hang over period, it can be useful to reduce the activation threshold, for example by a factor of 2/3. This creates increased sensitivity to voice and stability once a talk burst has commenced.
  • the noise in the above expression is N s determined using the out-of-beam power (or other frequency domain amplitude metric) 3 ⁇ 4' .
  • the values of ⁇ ⁇ , ⁇ ⁇ are not necessarily the same as for the echo update VAD 125.
  • This VAD is called a spatially- selective VAD and is shown as element 1021 in FIG. 10.
  • Y sens is set to be around expected microphone and system noise level, obtained by experiments on typical components.
  • ⁇ ⁇ , ⁇ ⁇ , ⁇ 5 ⁇ 5 , ⁇ thresh > BeamGainExp, and whether N3 ⁇ 4 or N3 ⁇ 4 g is used are tunable parameters, each tuned according to the function performed by the element in which an instantiation of the universal VAD is used. This is to enhance the voice quality while improving the suppression of undesired effects such as one or more of echoes, noise, and sounds from other than the speaker location.
  • Other uses for the VAD structures presented herein include the control of transmission or coding, level estimation, gain control and system power management. Wind activity detection
  • Some embodiments of the invention include a wind activity detector 1023 and wind activity detection step 1113 in the application of the gains, and in particular, in the postprocessing.
  • each of wind activity detector (WAD) 1023 and wind detecting step 1113 operates to detect the presence of corrupting wind influences in the plurality of inputs, e.g., microphone inputs, e.g., two microphone inputs.
  • the element 1023 and step 1113 determine an estimate of wind activity.
  • Any wind activity detector and wind detection method can be used in system and method embodiments of the invention.
  • the inventors chose to use the wind detector and wind detection method described in the Wind Detection/Suppression Application referenced in the "RELATED PATENT APPLICATIONS" Section herein above.
  • Some embodiments further include wind suppression. Wind suppression however is not discussed herein, but rather in the related Wind Detection/Suppression Application.
  • wind detector 1023 uses an algorithmic combination of
  • multiple features including spatial features to increase the specificity of the detection and reduce the occurrence of "false alarms” that would otherwise be caused by transient bursts of sound common in voice and acoustic interferers as is common in prior art wind detection.
  • a wind activity detector 1023 and a wind activity detection method 1113 use the following determined features for wind detection:
  • Slope the spectral slope, e.g., in dB per decade, obtained, for example, using
  • RatioStd the standard deviation of the difference between instantaneous and expected values of the ratio spatial feature, e.g., in dB, e.g., in the bands from 200 to
  • CoherStd the standard deviation of the coherence spatial feature in the bands from
  • Coherence ' b (can also be used in the log domain for analysis)
  • B bands In one embodiment, only some of the B bands are used. In one embodiment, a number of bands, typically between 5 and 20, covering the frequency range from approximately 200 to 1500 Hz are used. Slope is the linear relationship between ⁇ 0 ⁇ og o(Power) and logjo (BandFrequency). RatioStd is the standard deviation of the Ratio expressed in dB
  • Coherence Std is the standard deviation of Coherence expressed in dB (51ogio ' R b ⁇ 2 R b2 ⁇ ⁇ ) across the set of
  • Slope is the spectral slope, obtained from the current frame of data
  • WindSlopeBias and WindSlope are constants empirically determined, e.g., from plots of the power, in one embodiment arriving at the values -5 and -20, to achieve a scaling of the SlopeContribution such that 0 corresponds to no wind, 1 represents a nominal wind, and values greater 1 indicating progressively higher wind activity.
  • RatioStd is obtained from the current frame of data and WindRatioStd is a constant empirically determined from Ratio data over time to achieve a scaling of RatioContribution with the values 0 and 1 representing the absence and nominal level of wind as above.
  • CoherStd is obtained from the current frame of data and WindCoherStd is a constant empirically determined from Coherence data over time to achieve a scaling of CoherContribution with the values 0 and 1 representing the absence and nominal level of wind as above.
  • the overall wind level is then computed as the product
  • This overall wind level is a continuous variable with a value of 1 representing a
  • sensitivity can be increased or decreased as required for different detection requirements to balance sensitivity and specificity as needed.
  • a small offset e.g., 0.1 in one embodiment, is subtracted to remove some residual.
  • the signal can be further processed with smoothing or scaling to achieve the indicator of wind required for different functions.
  • a 100ms decay filter is used.
  • multiplication is in some form equivalent to the "ANDing" function.
  • multiple detections are used based on each indicator, in the form of:
  • SlopeContributionInd AND RatioContributionInd AND CoherContributionInd are the wind activity indicators based on SlopeContribution, RatioContribution, and CoherContribution, respectively.
  • a filter may be used to filter the WindLevel signal issuing from the wind detector. Due to the nature of wind and aspects of the detection method, this value can vary rapidly.
  • WindDecay reflects a first order time constant such that if the WindLevel were to be calculated at an interval of T, WindDecay varies as exp( - ⁇ 0.100), resulting in a time constant of 100ms.
  • a suitable threshold for creating a binary indicator of wind activity would sensibly be in the range of 0.2 to 1.5. In one embodiment a value of 1.0 was used against FilteredWindLevel to create a single binary indicator of wind.
  • system 100 includes suppressor element 131 to apply the (overall, post-processed) gain in B bands to simultaneously suppress noise, out- of-location signals, and in some embodiments, echoes from the banded mixed-down signal 108.
  • step 227 includes simultaneously suppressing noise, out-of- location signals, and in some embodiments suppressing echoes from the banded mixed-down signal by applying the (overall, post-processed) gain in B bands.
  • G n ⁇ w b ', n - G b '
  • w3 ⁇ 4 n represents an overlapping interpolation window.
  • the interpolation window is a raised cosine.
  • another widow such as a shape preserving spline, or other band-limited interpolation function is used.
  • ⁇ 1 ⁇ 4y n ⁇ 0 for all n.
  • the output syntheses process of step 229 is, in the case that the output is in the form of time samples, a conventional overlap add and inverse transform step, carried out, e.g., by output synthesizer/transformer 133.
  • step 229 The output remapping process of step 229 is, in the case that the output is in the
  • a remapper as needed for the following step, and carried out, e.g., by output remapper 133.
  • output remapper 133 In some embodiments, only time domain samples are output, in others only remapped frequency domain output is generated, while in yet other embodiments, both time domain output and remapped frequency domain output is generated. See FIGS. 3D and 3E.
  • a processing apparatus including a processing system
  • FIG. 16 shows a simplified block diagram of one processing apparatus embodiment 1600 for processing a plurality of audio inputs 101, e.g., from microphones (not shown) and one or more reference signals 102, e.g., from one or more loudspeakers (not shown) or from the feed(s) to such loudspeaker(s).
  • the processing apparatus 1600 is to generate audio output 135 that has been modified by suppressing, in one embodiment noise and out-of- location signals, and in another embodiment also echoes as specified in accordance to one or more features of the present invention.
  • the apparatus for example, can implement the system shown in FIG. 1 , and any alternates thereof, and can carry out, when operating, the method of FIG. 2 including any variations of the method described herein.
  • Such an apparatus may be included, for example, in a headphone set such as a Bluetooth headset.
  • the audio inputs 101 , the reference input(s) 102 and the audio output 135 are assumed to be in the form of frames of M samples of sampled data.
  • a digitizer including an analog-to-digital converter and quantizer would be present.
  • a de- quantizer and a digital-to-analog converter would be present.
  • FIG. 16 includes a processing system 1603 that is configured in operation to carry out the suppression methods described herein.
  • the processing system 1603 includes at least one processor 1605, which can be the processing unit(s) of a digital signal processing device, or a CPU of a more general purpose processing device.
  • the processing system 1603 also includes a storage subsystem 1607 typically including one or more memory elements.
  • the elements of the processing system are coupled, e.g., by a bus subsystem or some other interconnection mechanism not shown in FIG. 16. Some of the elements of processing system 1603 may be integrated into a single circuit, using techniques commonly known to one skilled in the art.
  • the storage subsystem 1607 includes instructions 1611 that when executed by the processor(s) 1605, cause carrying out of the methods described herein.
  • the storage subsystem 1607 is configured to store one or more tuning parameters 1613 that can be used to vary some of the processing steps carried out by the processing system 1603.
  • the system shown in FIG. 16 can be incorporated in a specialized device such as a headset, e.g., a wireless Bluetooth headset.
  • the system also can be part of a general purpose computer, e.g., a personal computer configured to process audio signals.
  • a suppression system embodiments and suppression method embodiments have been presented. The inventors have noted that it is possible to eliminate significant parts of the target signal without any perceptual distortion. The inventors note that the human brain is rather proficient at error correcting (particularly on voice) and thus many minor distortions in the form of unnecessary or unavoidable spectral suppression would still lead to perceptually pleasing results.
  • processing refers to the action and/or processes of a computer or computing system, or similar electronic computing device, that manipulate and/or transform data represented as physical, such as electronic, quantities into other data similarly represented as physical quantities.
  • processor may refer to any device or portion of a device that processes electronic data, e.g., from registers and/or memory to transform that electronic data into other electronic data that, e.g., may be stored in registers and/or memory.
  • a "computer” or a “computing machine” or a “computing platform” may include one or more processors.
  • the methodologies described herein are, in some embodiments, performable by one or more processors that accept logic, e.g., instructions encoded on one or more computer- readable media. When executed by one or more of the processors, the instructions cause carrying out at least one of the methods described herein. Any processor capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken is included. Thus, one example is a typical processing system that includes one or more processors. Each processor may include one or more of a CPU or similar element, a graphics processing unit (GPU), field-programmable gate array, application-specific integrated circuit, and/or a programmable DSP unit.
  • GPU graphics processing unit
  • DSP programmable DSP unit
  • the processing system further includes a storage subsystem with at least one storage medium, which may include memory embedded in a semiconductor device, or a separate memory subsystem including main RAM and/or a static RAM, and/or ROM, and also cache memory.
  • the storage subsystem may further include one or more other storage devices, such as magnetic and/or optical and/or further solid state storage devices.
  • a bus subsystem may be included for communicating between the components.
  • the processing system further may be a distributed processing system with processors coupled by a network, e.g., via network interface devices or wireless network interface devices.
  • the processing system requires a display, such a display may be included, e.g., a liquid crystal display (LCD), organic light emitting display (OLED), or a cathode ray tube (CRT) display.
  • a display e.g., a liquid crystal display (LCD), organic light emitting display (OLED), or a cathode ray tube (CRT) display.
  • the processing system also includes an input device such as one or more of an alphanumeric input unit such as a keyboard, a pointing control device such as a mouse, and so forth.
  • the term storage device, storage subsystem, or memory unit as used herein, if clear from the context and unless explicitly stated otherwise, also encompasses a storage system such as a disk drive unit.
  • the processing system in some configurations may include a sound output device, and a network interface device.
  • a non-transitory computer-readable medium is configured with, e.g., encoded with instructions, e.g., logic that when executed by one or more processors of a processing system such as a digital signal processing device or subsystem that includes at least one processor element and a storage subsystem, cause carrying out a method as described herein. Some embodiments are in the form of the logic itself.
  • a non-transitory computer-readable medium is any computer-readable medium that is statutory subject matter under the patent laws applicable to this disclosure, including Section 101 of Title 35 of the United States Code.
  • a non-transitory computer-readable medium is for example any computer-readable medium that is not specifically a transitory propagated signal or a transitory carrier wave or some other transitory transmission medium.
  • the term "non- transitory computer-readable medium” thus covers any tangible computer-readable storage medium.
  • the storage subsystem thus includes a computer-readable storage medium that is configured with, e.g., encoded with instructions, e.g., logic, e.g., software that when executed by one or more processors, causes carrying out one or more of the method steps described herein.
  • Non-transitory computer-readable media include any tangible computer- readable storage media and may take many forms including non- volatile storage media and volatile storage media.
  • Non- volatile storage media include, for example, static RAM, optical disks, magnetic disks, and magneto-optical disks.
  • Volatile storage media includes dynamic memory, such as main memory in a processing system, and hardware registers in a processing system.
  • the computer-readable medium is shown in an example embodiment to be a single medium, the term “medium” should be taken to include a single medium or multiple media (e.g., several memories, a centralized or distributed database, and/or associated caches and servers) that store the one or more sets of instructions.
  • a non-transitory computer-readable medium e.g., a computer-readable storage medium may form a computer program product, or be included in a computer program product.
  • the one or more processors operate as a standalone
  • processors may operate in the capacity of a server or a client machine in server-client network environment, or as a peer machine in a peer-to-peer or distributed network environment.
  • processing system encompasses all such possibilities, unless explicitly excluded herein.
  • the one or more processors may form a personal computer (PC), a media playback device, a headset device, a hands-free communication device, a tablet PC, a set-top box (STB), a Personal Digital Assistant (PDA), a game machine, a cellular telephone, a Web appliance, a network router, switch or bridge, or any machine capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by that machine.
  • PC personal computer
  • PDA Personal Digital Assistant
  • game machine a cellular telephone
  • Web appliance a network router, switch or bridge, or any machine capable of executing a set of instructions (sequential or otherwise) that specify actions to be taken by that machine.
  • embodiments of the present invention may be embodied as a method, an apparatus such as a special purpose apparatus, an apparatus such as a data processing system, logic, e.g., embodied in a non-transitory computer-readable medium, or a computer-readable medium that is encoded with instructions, e.g., a computer-readable storage medium configured as a computer program product.
  • the computer-readable medium is configured with a set of instructions that when executed by one or more processors cause carrying out method steps.
  • aspects of the present invention may take the form of a method, an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects.
  • the present invention may take the form of program logic, e.g., a computer program on a computer-readable storage medium, or the computer-readable storage medium configured with computer-readable program code, e.g., a computer program product.
  • an element described herein of an apparatus embodiment is an example of a means for carrying out the function performed by the element for the purpose of carrying out the invention.
  • the invention is not limited to use of power, i.e., the weighted sum of the squares of the frequency coefficient amplitudes, and can be modified to accommodate any metric of the amplitude.
  • any one of the terms comprising, comprised of or which comprises is an open term that means including at least the elements/features that follow, but not excluding others.
  • the term comprising, when used in the claims should not be interpreted as being limitative to the means or elements or steps listed thereafter.
  • the scope of the expression a device comprising element_A and element_B should not be limited to devices consisting of only elements element_A and element_B.
  • Any one of the terms including or which includes or that includes as used herein is also an open term that also means including at least the elements/features that follow the term, but not excluding others. Thus, including is synonymous with and means comprising.
  • Coupled when used in the claims, should not be interpreted as being limitative to direct connections only.
  • the terms “coupled” and “connected,” along with their derivatives, may be used. It should be understood that these terms are not intended as synonyms for each other, but may be.
  • the scope of the expression “a device A coupled to a device B” should not be limited to devices or systems wherein an input or output of device A is directly connected to an output or input of device B. It means that there exists a path between device A and device B which may be a path including other devices or means in between.
  • coupled to does not imply direction.
  • a device A is coupled to a device B
  • a device B is coupled to a device A
  • Coupled may mean that two or more elements are either in direct physical or electrical contact, or that two or more elements are not in direct contact with each other but yet still co-operate or interact with each other.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Quality & Reliability (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Circuit For Audible Band Transducer (AREA)
  • Soundproofing, Sound Blocking, And Sound Damping (AREA)
  • Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)

Abstract

L'invention concerne un système, un procédé, une logique mise en œuvre dans un support lisible par ordinateur, et un support lisible par ordinateur comprenant des instructions qui lorsqu'elles sont exécutées mettent en œuvre un procédé. Le procédé comprend: (a) plusieurs signaux d'entrée, par ex., des signaux provenant de plusieurs microphones séparés spatialement; et, pour supprimer l'écho, (b) un ou plusieurs signaux de référence, par ex., des signaux provenant de, ou produits par un ou plusieurs haut-parleurs et pouvant produire des échos. Le procédé comprend des signaux d'entrée et un ou plusieurs signaux de référence destinés à mettre en œuvre de manière intégrée, simultanément la suppression du bruit et la suppression du signal hors emplacement, et dans certaines versions, la suppression de l'écho.
PCT/US2012/024370 2011-02-10 2012-02-08 Suppression de bruit combinée et signaux hors emplacement WO2012109384A1 (fr)

Priority Applications (4)

Application Number Priority Date Filing Date Title
JP2013553528A JP6002690B2 (ja) 2011-02-10 2012-02-08 オーディオ入力信号処理システム
CN201280008266.XA CN103348408B (zh) 2011-02-10 2012-02-08 噪声和位置外信号的组合抑制方法和系统
EP12707412.8A EP2673777B1 (fr) 2011-02-10 2012-02-08 Suppression de bruit combinée et signaux hors emplacement
US13/964,037 US9173025B2 (en) 2012-02-08 2013-08-09 Combined suppression of noise, echo, and out-of-location signals

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US201161441611P 2011-02-10 2011-02-10
US61/441,611 2011-02-10

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US13/964,037 Continuation US9173025B2 (en) 2012-02-08 2013-08-09 Combined suppression of noise, echo, and out-of-location signals

Publications (1)

Publication Number Publication Date
WO2012109384A1 true WO2012109384A1 (fr) 2012-08-16

Family

ID=45809594

Family Applications (2)

Application Number Title Priority Date Filing Date
PCT/US2012/024372 WO2012109385A1 (fr) 2011-02-10 2012-02-08 Post-traitement comprenant le filtrage médian de gains de suppression de bruit
PCT/US2012/024370 WO2012109384A1 (fr) 2011-02-10 2012-02-08 Suppression de bruit combinée et signaux hors emplacement

Family Applications Before (1)

Application Number Title Priority Date Filing Date
PCT/US2012/024372 WO2012109385A1 (fr) 2011-02-10 2012-02-08 Post-traitement comprenant le filtrage médian de gains de suppression de bruit

Country Status (4)

Country Link
EP (2) EP2673777B1 (fr)
JP (1) JP6002690B2 (fr)
CN (2) CN103354937B (fr)
WO (2) WO2012109385A1 (fr)

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013142661A1 (fr) 2012-03-23 2013-09-26 Dolby Laboratories Licensing Corporation Post-traitement des gains pour l'amélioration du signal
US8804977B2 (en) 2011-03-18 2014-08-12 Dolby Laboratories Licensing Corporation Nonlinear reference signal processing for echo suppression
WO2014143582A1 (fr) 2013-03-14 2014-09-18 Dolby Laboratories Licensing Corporation Bruit de confort spatial
JP2015037239A (ja) * 2013-08-13 2015-02-23 日本電信電話株式会社 残響抑圧装置とその方法と、プログラムとその記録媒体
TWI498884B (zh) * 2013-09-09 2015-09-01 Pegatron Corp 具有過濾背景音功能的電子裝置及其方法
US9313597B2 (en) 2011-02-10 2016-04-12 Dolby Laboratories Licensing Corporation System and method for wind detection and suppression
US9538286B2 (en) 2011-02-10 2017-01-03 Dolby International Ab Spatial adaptation in multi-microphone sound capture
US10021501B2 (en) 2013-09-27 2018-07-10 Fraunhofer-Gesellschaft Zur Foerderung Der Angewandten Forschung E.V. Concept for generating a downmix signal
US10079941B2 (en) 2014-07-07 2018-09-18 Dolby Laboratories Licensing Corporation Audio capture and render device having a visual display and user interface for use for audio conferencing
CN111145770A (zh) * 2018-11-02 2020-05-12 北京微播视界科技有限公司 音频处理方法和装置
GB2603548A (en) * 2021-02-04 2022-08-10 Neatframe Ltd Audio processing
WO2022167553A1 (fr) * 2021-02-04 2022-08-11 Neatframe Limited Traitement audio
WO2022189481A1 (fr) * 2021-03-11 2022-09-15 Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. Décorrélateur audio, système de traitement et procédé de décorrélation d'un signal audio
US20220369030A1 (en) * 2021-05-17 2022-11-17 Apple Inc. Spatially informed acoustic echo cancelation
US11513205B2 (en) 2017-10-30 2022-11-29 The Research Foundation For The State University Of New York System and method associated with user authentication based on an acoustic-based echo-signature
EP4198976A1 (fr) * 2021-12-17 2023-06-21 GN Audio A/S Système de suppression de bruit du vent
US11880407B2 (en) 2015-06-30 2024-01-23 Fraunhofer-Gesellschaft Zur Foerderung Der Angewandten Forschung E.V. Method and device for generating a database of noise

Families Citing this family (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB201309777D0 (en) * 2013-05-31 2013-07-17 Microsoft Corp Echo suppression
GB201309773D0 (en) 2013-05-31 2013-07-17 Microsoft Corp Echo removal
GB201309779D0 (en) 2013-05-31 2013-07-17 Microsoft Corp Echo removal
GB201309771D0 (en) 2013-05-31 2013-07-17 Microsoft Corp Echo removal
CN104681034A (zh) 2013-11-27 2015-06-03 杜比实验室特许公司 音频信号处理
CN105794190B (zh) * 2013-12-12 2019-09-20 皇家飞利浦有限公司 一种音频回声抑制器及音频回声抑制方法
CN104200817B (zh) * 2014-07-31 2017-07-28 广东美的制冷设备有限公司 语音控制方法和系统
JP6195548B2 (ja) * 2014-08-19 2017-09-13 日本電信電話株式会社 信号解析装置、方法、及びプログラム
US9953661B2 (en) * 2014-09-26 2018-04-24 Cirrus Logic Inc. Neural network voice activity detection employing running range normalization
WO2016112968A1 (fr) * 2015-01-14 2016-07-21 Widex A/S Procédé pour faire fonctionner un système d'aide auditive, et système d'aide auditive
US9401158B1 (en) * 2015-09-14 2016-07-26 Knowles Electronics, Llc Microphone signal fusion
CN107493247B (zh) * 2016-06-13 2021-10-22 中兴通讯股份有限公司 一种自适应均衡方法、装置及均衡器
US9906859B1 (en) * 2016-09-30 2018-02-27 Bose Corporation Noise estimation for dynamic sound adjustment
CN106782591B (zh) * 2016-12-26 2021-02-19 惠州Tcl移动通信有限公司 一种在背景噪音下提高语音识别率的装置及其方法
CN107886965B (zh) * 2017-11-28 2021-04-20 游密科技(深圳)有限公司 游戏背景音的回声消除方法
CN108447500B (zh) * 2018-04-27 2020-08-18 深圳市沃特沃德股份有限公司 语音增强的方法与装置
US10847178B2 (en) * 2018-05-18 2020-11-24 Sonos, Inc. Linear filtering for noise-suppressed speech detection
WO2020023856A1 (fr) * 2018-07-27 2020-01-30 Dolby Laboratories Licensing Corporation Insertion d'intervalle forcé pour écoute omniprésente
CN109949820B (zh) * 2019-03-07 2020-05-08 出门问问信息科技有限公司 一种语音信号处理方法、装置及系统
CN110097884B (zh) * 2019-06-11 2022-05-17 大众问问(北京)信息科技有限公司 一种语音交互方法和装置
CN110136739B (zh) * 2019-07-03 2021-04-09 四川大学 语音信号中冲激性噪声的去除方法
CN112420073B (zh) * 2020-10-12 2024-04-16 北京百度网讯科技有限公司 语音信号处理方法、装置、电子设备和存储介质
CN112235693B (zh) * 2020-11-04 2021-12-21 北京声智科技有限公司 麦克风信号处理方法、装置、设备及计算机可读存储介质
CN113689878A (zh) * 2021-07-26 2021-11-23 浙江大华技术股份有限公司 回声消除方法、回声消除装置及计算机可读存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040016964A1 (en) 2002-07-26 2004-01-29 Ji-Young Kim Semiconductor device with self-aligned junction contact hole and method of fabricating the same
WO2004111994A2 (fr) 2003-05-28 2004-12-23 Dolby Laboratories Licensing Corporation Procede, appareil et programme informatique pour le calcul et le reglage de la force sonore perçue d'un signal sonore
US20080288219A1 (en) * 2007-05-17 2008-11-20 Microsoft Corporation Sensor array beamformer post-processor
US20080317259A1 (en) * 2006-05-09 2008-12-25 Fortemedia, Inc. Method and apparatus for noise suppression in a small array microphone system
US20100241426A1 (en) * 2009-03-23 2010-09-23 Vimicro Electronics Corporation Method and system for noise reduction

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2913105B2 (ja) * 1989-03-10 1999-06-28 日本電信電話株式会社 音響信号検出方法
DE4405723A1 (de) * 1994-02-23 1995-08-24 Daimler Benz Ag Verfahren zur Geräuschreduktion eines gestörten Sprachsignals
JP3484801B2 (ja) * 1995-02-17 2004-01-06 ソニー株式会社 音声信号の雑音低減方法及び装置
US5587998A (en) * 1995-03-03 1996-12-24 At&T Method and apparatus for reducing residual far-end echo in voice communication networks
US6351731B1 (en) * 1998-08-21 2002-02-26 Polycom, Inc. Adaptive filter featuring spectral gain smoothing and variable noise multiplier for noise reduction, and method therefor
JP3435357B2 (ja) * 1998-09-07 2003-08-11 日本電信電話株式会社 収音方法、その装置及びプログラム記録媒体
JP4247037B2 (ja) * 2003-01-29 2009-04-02 株式会社東芝 音声信号処理方法と装置及びプログラム
ATE413769T1 (de) * 2004-09-03 2008-11-15 Harman Becker Automotive Sys Sprachsignalverarbeitung für die gemeinsame adaptive reduktion von störgeräuschen und von akustischen echos
US8280730B2 (en) * 2005-05-25 2012-10-02 Motorola Mobility Llc Method and apparatus of increasing speech intelligibility in noisy environments
CN101263734B (zh) * 2005-09-02 2012-01-25 丰田自动车株式会社 麦克风阵列用后置滤波器
WO2007049643A1 (fr) * 2005-10-26 2007-05-03 Nec Corporation Procede et dispositif d'annulation d'echo
JP4850191B2 (ja) * 2008-01-16 2012-01-11 富士通株式会社 自動音量制御装置及びそれを用いた音声通信装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040016964A1 (en) 2002-07-26 2004-01-29 Ji-Young Kim Semiconductor device with self-aligned junction contact hole and method of fabricating the same
WO2004111994A2 (fr) 2003-05-28 2004-12-23 Dolby Laboratories Licensing Corporation Procede, appareil et programme informatique pour le calcul et le reglage de la force sonore perçue d'un signal sonore
US20080317259A1 (en) * 2006-05-09 2008-12-25 Fortemedia, Inc. Method and apparatus for noise suppression in a small array microphone system
US20080288219A1 (en) * 2007-05-17 2008-11-20 Microsoft Corporation Sensor array beamformer post-processor
US20100241426A1 (en) * 2009-03-23 2010-09-23 Vimicro Electronics Corporation Method and system for noise reduction

Non-Patent Citations (10)

* Cited by examiner, † Cited by third party
Title
AUDONE, B.; M. BUZZO MARGARI: "The use of MUSIC algorithm to characterize emissive sources", ELECTROMAGNETIC COMPATIBILITY, IEEE TRANSACTIONS ON, vol. 43, no. 4, 2001, pages 688 - 693, XP011019772
B. C. J. MOORE; B. GLASBERG; T. BAER: "A Model for the Prediction of Thresholds, Loudness, and Partial Loudness", J. OF THE AUDIO ENGINEERING SOCIETY (AES, vol. 45, no. 4, April 1997 (1997-04-01), pages 224 - 240, XP000700661
JOHNSON, D.H.; D.E. DUDGEON: "Array Signal Processing: Concepts and Techniques", 1993, PRENTICE HALL
R. MARTIN: "Spectral Subtraction Based on Minimum Statistics", PROC. EURO. SIGNAL PROCESSING CONF. (EUSIPCO, 1994, pages 1182 - 1185, XP009026531
ROY, R.; A. PAULRAJ; T. KAILATH: "ESPRIT -A subspace rotation approach to estimation of parameters of cisoids in noise", IEEE TRANS. ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, vol. 34, no. 5, 1986, pages 1340 - 1342, XP000570836, DOI: doi:10.1109/TASSP.1986.1164935
STOICA, P.; A. NEHORAI: "MUSIC, maximum likelihood, and Cramer-Rao bound", IEEE TRANS. ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, vol. 37, no. 5, 1989, pages 720 - 741
TERHARDT, E., CALCULATING VIRTUAL PITCH. HEARING RESEARCH, vol. 1, 1979, pages 155 - 182
VAN TREES, H.L.: "Detection, estimation, and modulation theory: {IV} Optimum Array Processing", 2002, WILEY
WAX, M.; I. ZISKIND: "On unique localization of multiple sources by passive sensor arrays", IEEE TRANS. ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, vol. 37, no. 7, 1989, pages 996 - 1000, XP000037319, DOI: doi:10.1109/29.32277
WITTKOP T ET AL: "SPEECH PROCESSING FOR HEARING AIDS: NOISE REDUCTION MOTIVATED BY MODELS OF BINAURAL INTERACTION", ACTA ACUSTICA, EDITIONS DE PHYSIQUE. LES ULIS CEDEX, FR, vol. 83, no. 4, 1 January 1997 (1997-01-01), pages 684 - 699, XP000884158 *

Cited By (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9313597B2 (en) 2011-02-10 2016-04-12 Dolby Laboratories Licensing Corporation System and method for wind detection and suppression
US10154342B2 (en) 2011-02-10 2018-12-11 Dolby International Ab Spatial adaptation in multi-microphone sound capture
US9761214B2 (en) 2011-02-10 2017-09-12 Dolby Laboratories Licensing Corporation System and method for wind detection and suppression
US9538286B2 (en) 2011-02-10 2017-01-03 Dolby International Ab Spatial adaptation in multi-microphone sound capture
US8804977B2 (en) 2011-03-18 2014-08-12 Dolby Laboratories Licensing Corporation Nonlinear reference signal processing for echo suppression
US10311891B2 (en) 2012-03-23 2019-06-04 Dolby Laboratories Licensing Corporation Post-processing gains for signal enhancement
US11308976B2 (en) 2012-03-23 2022-04-19 Dolby Laboratories Licensing Corporation Post-processing gains for signal enhancement
US9584087B2 (en) 2012-03-23 2017-02-28 Dolby Laboratories Licensing Corporation Post-processing gains for signal enhancement
WO2013142661A1 (fr) 2012-03-23 2013-09-26 Dolby Laboratories Licensing Corporation Post-traitement des gains pour l'amélioration du signal
US11694711B2 (en) 2012-03-23 2023-07-04 Dolby Laboratories Licensing Corporation Post-processing gains for signal enhancement
US10902865B2 (en) 2012-03-23 2021-01-26 Dolby Laboratories Licensing Corporation Post-processing gains for signal enhancement
WO2014143582A1 (fr) 2013-03-14 2014-09-18 Dolby Laboratories Licensing Corporation Bruit de confort spatial
US10224046B2 (en) 2013-03-14 2019-03-05 Dolby Laboratories Licensing Corporation Spatial comfort noise
JP2015037239A (ja) * 2013-08-13 2015-02-23 日本電信電話株式会社 残響抑圧装置とその方法と、プログラムとその記録媒体
TWI498884B (zh) * 2013-09-09 2015-09-01 Pegatron Corp 具有過濾背景音功能的電子裝置及其方法
US10021501B2 (en) 2013-09-27 2018-07-10 Fraunhofer-Gesellschaft Zur Foerderung Der Angewandten Forschung E.V. Concept for generating a downmix signal
RU2661310C2 (ru) * 2013-09-27 2018-07-13 Фраунхофер-Гезелльшафт Цур Фердерунг Дер Ангевандтен Форшунг Е.Ф. Концепция генерирования сигнала понижающего микширования
US10079941B2 (en) 2014-07-07 2018-09-18 Dolby Laboratories Licensing Corporation Audio capture and render device having a visual display and user interface for use for audio conferencing
US11880407B2 (en) 2015-06-30 2024-01-23 Fraunhofer-Gesellschaft Zur Foerderung Der Angewandten Forschung E.V. Method and device for generating a database of noise
US11513205B2 (en) 2017-10-30 2022-11-29 The Research Foundation For The State University Of New York System and method associated with user authentication based on an acoustic-based echo-signature
CN111145770B (zh) * 2018-11-02 2022-11-22 北京微播视界科技有限公司 音频处理方法和装置
CN111145770A (zh) * 2018-11-02 2020-05-12 北京微播视界科技有限公司 音频处理方法和装置
GB2603548A (en) * 2021-02-04 2022-08-10 Neatframe Ltd Audio processing
WO2022167553A1 (fr) * 2021-02-04 2022-08-11 Neatframe Limited Traitement audio
WO2022189481A1 (fr) * 2021-03-11 2022-09-15 Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. Décorrélateur audio, système de traitement et procédé de décorrélation d'un signal audio
US20220369030A1 (en) * 2021-05-17 2022-11-17 Apple Inc. Spatially informed acoustic echo cancelation
US11849291B2 (en) * 2021-05-17 2023-12-19 Apple Inc. Spatially informed acoustic echo cancelation
EP4198976A1 (fr) * 2021-12-17 2023-06-21 GN Audio A/S Système de suppression de bruit du vent

Also Published As

Publication number Publication date
CN103348408B (zh) 2015-11-25
EP2673777B1 (fr) 2018-12-26
EP2673777A1 (fr) 2013-12-18
JP2014510452A (ja) 2014-04-24
CN103348408A (zh) 2013-10-09
EP2673778A1 (fr) 2013-12-18
CN103354937A (zh) 2013-10-16
WO2012109385A1 (fr) 2012-08-16
CN103354937B (zh) 2015-07-29
JP6002690B2 (ja) 2016-10-05
EP2673778B1 (fr) 2018-10-10

Similar Documents

Publication Publication Date Title
US9173025B2 (en) Combined suppression of noise, echo, and out-of-location signals
EP2673777B1 (fr) Suppression de bruit combinée et signaux hors emplacement
JP6014259B2 (ja) ノイズ削減利得の百分位数フィルタリング
Doclo et al. Frequency-domain criterion for the speech distortion weighted multichannel Wiener filter for robust noise reduction
US8521530B1 (en) System and method for enhancing a monaural audio signal
US8654990B2 (en) Multiple microphone based directional sound filter
US7366662B2 (en) Separation of target acoustic signals in a multi-transducer arrangement
US8712076B2 (en) Post-processing including median filtering of noise suppression gains
RU2760097C2 (ru) Способ и устройство для захвата аудиоинформации с использованием формирования диаграммы направленности
US9532149B2 (en) Method of signal processing in a hearing aid system and a hearing aid system
EP3275208B1 (fr) Mélange de sous-bande de multiples microphones
TW201142829A (en) Adaptive noise reduction using level cues
US10262673B2 (en) Soft-talk audio capture for mobile devices
EP3120355A2 (fr) Suppression de bruit
GB2453118A (en) Generating a speech audio signal from multiple microphones with suppressed wind noise
US20200286501A1 (en) Apparatus and a method for signal enhancement
JP2005514668A (ja) スペクトル出力比依存のプロセッサを有する音声向上システム
Kodrasi et al. Curvature-based optimization of the trade-off parameter in the speech distortion weighted multichannel wiener filter
Ahmed et al. Comparison of LMS and NLMS algorithm with the using of 4 Linear Microphone Array for Speech Enhancement.
Xiong et al. A study on joint beamforming and spectral enhancement for robust speech recognition in reverberant environments
Gustafsson et al. Dual-Microphone Spectral Subtraction
Nahma A Study into Speech Enhancement Techniques in Adverse Environment

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 12707412

Country of ref document: EP

Kind code of ref document: A1

DPE1 Request for preliminary examination filed after expiration of 19th month from priority date (pct application filed from 20040101)
WWE Wipo information: entry into national phase

Ref document number: 2012707412

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2013553528

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE