JP2008070878A - Voice signal pre-processing device, voice signal processing device, voice signal pre-processing method and program for voice signal pre-processing - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To effectively reduce external noise from various vibration sources. <P>SOLUTION: The voice signal pre-processing device 10 including a first filter 20 for reducing noise and a second filter 22 for reducing noise, reduces a noise component by processing an electric signal including voice. Either of the first filter 20 or the second filter 22 is a coherent filter for reducing noise by coherent filtering, and the other is a non-coherent filter for reducing noise by non-coherent filtering, and the second filter 22 performs filtering of output of the first filter 20. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention preprocesses an audio-containing signal and reduces an external noise component included in the signal, an audio signal preprocessing device including the audio signal preprocessing device, an audio signal preprocessing method, and an audio signal The present invention relates to a preprocessing program. In particular, it is particularly suitable for processing a signal input to a voice recognition system or a telephone (for example, a hands-free telephone). Furthermore, the present invention is particularly suitable for use in a vehicle environment in which a relatively large amount of external noise interferes with the speech recognition process, thereby affecting the accuracy and reliability of speech recognition.

  There has been widespread interest in using voice for hands-free operation of in-vehicle electrical equipment and using hands-free telephones in the passenger compartment. While electronic voice recognition technology is used to recognize voice commands and other speech information collected by microphones, these voice recognition technologies are suitable for relatively “clean” voice signals that do not contain extraneous noise. . However, various external noise sources that severely degrade the quality of the collected audio signal exist in the vehicle environment, making speech recognition difficult or reducing the reliability of speech recognition. For example, external noise sources include road noise, engine noise, tire noise, wind noise, rain sound, sound from radio and music players, rattling and vibration in the vehicle, windshield wiper sound, For example, transient noise. It is difficult to filter these various extraneous noises related to speech recognition results without destroying many components of the desired speech signal. Furthermore, in order to be accepted by vehicle manufacturers and vehicle parts manufacturers, these technologies must exhibit considerable performance without cost and without adding a lot of new equipment and processing devices.

  The prior art presents various techniques for reducing these noises and improving the quality of audio components included in audio-containing signals.

  One of them is a method of using a microphone network or a beam forming technique to control the directivity of sound collection. For example, the beamforming technique can direct the directivity toward the driver's direction and position. However, noise from a specific direction in the vehicle compartment is rarely the main cause, so this alone can only reduce the external noise only a little, and adding multiple microphones can cause costs. And not very effective.

  As another method, two microphones are used, one is a first microphone that faces in a direction of collecting both voice and external noise, and the other is a second microphone that mainly faces in a direction of collecting external noise. The second microphone provides a noise reference signal that is input to a noise canceling filter that reduces noise from the audio-containing signal collected by the first microphone. However, this technique may be inappropriate depending on the placement of the microphone. In order to reduce the amount of sound collected by the second microphone, it is necessary to dispose the second microphone away from the first microphone. However, the longer the distance between the second microphone and the first microphone, the worse the performance of the second microphone as a reference signal for local noise (external noise) around the first microphone.

  An audio signal processing apparatus that receives an electric drive signal used directly to drive a vehicle-mounted audio system speaker is disclosed (see Patent Document 1). In Patent Document 1, this electric drive signal is an exact substitute for an external sound of an audio system to be reduced from a signal collected by a microphone (hereinafter referred to as a microphone signal), and is used as a noise reference signal to a noise cancellation filter. As provided, there is no need to add a sensor to obtain a noise reference signal.

Patent Document 2 discloses a continuous noise cancellation system that divides a microphone signal into a plurality of frequency bands and each band. In Patent Document 2, it is determined in principle whether the main external noise component is coherent or non-coherent. Either a coherent noise canceling filter or an adaptive non-coherent noise canceling filter is selectively performed in response to the determination. After filtering in an optimal manner for each band, a plurality of frequency bands are combined to reproduce the signal. Transient noise that makes filtering performance unstable is detected before the start of main filtering. The adaptive non-coherent noise canceling filter receives a signal from an external non-acoustic sensor such as a vibration sensor and performs spectral power filtering on the estimated transfer function between the microphone and the non-acoustic sensor.
JP-A-2-244099 JP 2006-276856 A

  In the audio signal processing device of Patent Document 1, this device cannot reduce external noise from various other vibration sources such as vibrations of the engine, roads, tires, wind, rain, vehicles, and the like. Further, even the noise cancellation system of Patent Document 2 is considered not sufficient because it can reduce external noise from various vibration sources.

  Therefore, the present invention provides an audio signal preprocessing device, an audio signal processing device, an audio signal preprocessing method, and an audio signal preprocessing program that can effectively reduce external noise from various vibration sources. is there.

  What solves the said subject is as follows.

(1) In an audio signal pre-processing device that includes a noise-reducing first filter and a noise-reducing second filter, and that processes an electric audio signal including audio to reduce noise components, the first filter or the first filter One of the two filters is a coherent filter that reduces the noise component by coherent filtering, and the other is a non-coherent filter that reduces the noise component by non-coherent filtering, and the second filter filters the output of the first filter An audio signal pre-processing apparatus.

(2) The audio signal preprocessing device according to (1), wherein the first filter is the coherent filter, and the second filter is the non-coherent filter.

(3) The coherent filter is configured to receive from a non-acoustic coherent noise sensor a first noise reference signal that is coherent with a noise component included in the electrical audio signal. Or the audio | voice signal pre-processing apparatus as described in (2).

(4) The non-coherent filter is configured to receive, from a non-acoustic non-coherent noise sensor, a second noise reference signal that is non-coherent with a noise component included in the electrical audio signal and related in spectrum power. The audio signal preprocessing device according to any one of (1) to (3) above, wherein

(5) The audio signal preprocessing device according to (3) or (4), wherein the non-acoustic coherent noise sensor is non-acoustically connected to an audio system of a vehicle.

(6) The audio signal preprocessing device according to (4) or (5), wherein the non-acoustic non-coherent noise sensor is configured to sense a vibration of a vehicle.

(7) The audio signal preprocessing device according to any one of (1) to (6), wherein the coherent filter includes a linear filter.

(8) The audio signal preprocessing device according to any one of (1) to (7), wherein the coherent filter includes a Wiener filter.

(9) Any one of (3) to (8), wherein the coherent filter includes a filter coefficient calculation unit that repeatedly calculates a filter coefficient based on the electrical audio signal and the first noise reference signal. The audio signal preprocessing device according to 1.

(10) The audio signal preprocessing device according to any one of (1) to (9), wherein the non-coherent filter includes a nonlinear filter.

(11) The audio signal preprocessing device according to any one of (1) to (10), wherein the non-coherent filter includes a spectrum-gain filter.

(12) The noncoherent filter further includes a filter coefficient calculation unit that calculates a filter coefficient based on the electrical audio signal and the second noise reference signal input to the noncoherent filter. 4) The audio signal preprocessing device according to any one of (11).

(13) Each of the first filter and the second filter has a plurality of sub-filters that filter the electrical audio signal in a plurality of corresponding frequency subbands. The audio signal preprocessing device according to any one of the above.

(14) The above (13), further comprising: an input unit that divides the electrical audio signal into the frequency subbands; and an output unit that synthesizes the electrical audio signal from the frequency subbands after filtering. The audio signal pre-processing device described.

(15) The first filter and the second filter process the electrical audio signal as a series of frames, and each frame processes the signal as a continuous frame of frames representing a signal in a window for a predetermined period. The audio signal preprocessing device according to any one of (1) to (14) above,

(16) The method according to (15), further comprising: an input unit that divides the electrical audio signal into the series of frames; and an output unit that synthesizes the electrical audio signal from the series of frames. Audio signal pre-processing device.

(17) The audio signal preprocessing device according to any one of (1) to (16),
A microphone that collects an acoustic signal and supplies the electrical audio signal as an input to the audio signal preprocessing device;
The non-acoustic coherent noise sensor that generates the first noise reference signal coherent with a first noise source that can be collected by the microphone, and supplies the first noise reference signal to the coherent filter;
Generating a second noise reference signal that is non-coherent with a second noise source that can be collected by the microphone and is related in spectrum power to the second noise source; An audio signal processing apparatus comprising: the non-acoustic non-coherent noise sensor supplied to a non-coherent filter.

(18) The audio signal processing device according to (17), wherein the non-acoustic non-coherent noise sensor is a vibration sensor or an accelerometer.

(19) In a speech preprocessing method for filtering an electrical speech signal including speech and reducing a noise component, a first filtering step of performing either one of coherent filtering or non-coherent filtering on the electrical speech signal by a first filter. When,
And a second filtering step of filtering the output based on the first filter by the second filter, which is the other filtering different from the first filter.

(20) The audio signal preprocessing method according to (19), wherein the first filter is the coherent filter, and the second filter is the non-coherent filter.

(21) A program for audio signal preprocessing, which causes a computer to execute the method according to (19) or (20).

  Moreover, it is preferable that the audio signal preprocessing device of the present invention is mounted on an integrated circuit. The audio signal preprocessing device is preferably applied to a vehicle for reducing external vehicle noise. Moreover, it is preferable to further have a voice recognition unit to which the output of the voice signal preprocessing device is input.

  According to the audio signal pre-processing device of the first aspect, since it is sequentially filtered by both the coherent filter and the non-coherent filter, various external noise components can be effectively filtered.

  According to the speech signal preprocessing device of claim 2, the coherent filtering can be optimally reduced by performing the coherent filtering before the non-coherent filtering, and further, the non-coherent filtering is performed by a component that can be removed or at least reduced by the coherent filtering. It can be avoided that the coherent filtering is biased.

  According to the speech signal preprocessing devices of claims 3 to 6, since the noise reference signal can be acquired non-acousticly, a speech segment detector is not necessary.

  According to the audio signal processing device of the seventeenth aspect, since filtering is sequentially performed by both the coherent filter and the non-coherent filter, various external noise components can be effectively filtered.

  According to the audio signal preprocessing method of the nineteenth aspect, since filtering is sequentially performed by both the coherent filter and the non-coherent filter, various external noise components can be effectively filtered.

  According to the speech signal preprocessing method of claim 20, the coherent filtering can be optimally reduced by performing the coherent filtering before the non-coherent filtering, and further, the non-coherent filtering can be performed by a component that can be removed or at least reduced by the coherent filtering. It can be avoided that the coherent filtering is biased.

  According to the audio signal preprocessing program of the twenty-first aspect, since filtering is sequentially performed by both the coherent filter and the non-coherent filter, various external noise components can be effectively filtered.

  First, an outline of the present invention will be described.

  Embodiment 1 of the present invention relates to effective filtering of a plurality of noises.

  The audio signal pre-processing device (audio signal processing device) of the present invention provides a filter having a first filter 20 and a second filter 22 that are continuous with each other. One of the first and second filters 20 and 22 is a coherent filter that reduces a coherent external noise component, and the other is a non-coherent filter that reduces a non-coherent external noise component. The first filter 20 and the second filter 22 are continuously arranged. With such a combination of continuous filters, the audio signal is sequentially filtered by both the coherent filter and the non-coherent filter, so that various external noise components can be effectively filtered. It is particularly suitable for vehicle use.

  In one form, the first filter 20 is a coherent filter and the second filter 22 is a non-coherent filter. By performing the coherent filtering before the non-coherent filtering, the coherent component can be optimally reduced, and further, the non-coherent filtering can be avoided from being distorted by a component that can be removed or at least reduced by the coherent filtering.

  The coherent filter and the non-coherent filter receive noise reference signals from respective noise reference sources. Each noise reference source is a non-acoustic noise reference source. Non-acoustic means that the noise reference source does not directly detect vibrations in the air. However, the non-acoustic noise reference source detects vibrations in the audible frequency range at a location in the vehicle and generates a signal indicative of the components in the audible frequency range.

  The second embodiment of the present invention relates to spectrum power (spectrum-gain) filtering. The filter used here is an iterative filter. This filter produces a calibration gain associated with an estimate of the magnitude of the transfer function between the noise reference source and the microphone. The calibration gain is a single value / signal or a spectral value / signal.

  The audio signal preprocessing apparatus (audio signal processing apparatus) of the present invention includes a limiting unit that monitors and limits the maximum allowable change rate of the calibration gain. The second embodiment is based on the recognition that the magnitude of the transfer function changes but changes relatively slowly (that is, the rate of change is smaller than a certain limit value). If the generated calibration gain begins to fluctuate at a fast rate of change, this indicates that the microphone signal has become unstable due to other components not related to the noise reference source, such as speech or transient noise.

  As described above, since the limiting unit automatically prevents the disturbance caused by the transient noise or the voice, there is no need to provide a dedicated transient noise detector or a dedicated voice section detector.

  Preferably, the upper limit value of the change threshold value may be set sufficiently large within an expected change rate range so that the calibration gain can cope with a change in the size of the transfer function. Furthermore, the lower limit value of the change threshold value may be set sufficiently small so that the calibration gain is not disturbed by a rapid fluctuation of a component not related to the transfer function.

  The calibration gain is periodically generated, and the newly generated calibration gain is compared with the comparison (reference) value. The comparison value is obtained from one or more previous values of the calibration gain. If the difference between the new calibration gain and the comparison value exceeds a certain threshold, the new calibration gain is replaced with the comparison value. This threshold is a predetermined ratio of the comparison value, for example, about 20% of the comparison value.

  The calibration gain has a variable magnification. To account for speed related noise, the variable magnification may depend on the vehicle speed.

  Embodiments described herein relate generally to a technique for reducing vehicle noise from a microphone signal. According to this, a clearer voice signal can be obtained, and the voice can be recognized more effectively for an in-vehicle voice recognition system or a hands-free phone. According to the first embodiment, coherent filtering and non-coherent filtering are sequentially performed. Each filtering is performed based on a corresponding non-acoustic noise reference source. Coherent filtering may be performed before non-coherent filtering. According to the second embodiment, spectrum-gain filtering is performed by limiting the maximum allowable change rate of the calibration gain related to the estimated value of the transfer function between the noise reference source and the microphone. In this way, by limiting the maximum allowable change rate of the calibration gain, it is possible to automatically prevent transient noise and turbulence due to speech, resulting in the need for a speech interval detector as in the past. Disappears.

  Although the above aspects of the present invention can be implemented separately, these different aspects may be combined.

  Next, the present invention will be specifically described.

(Embodiment 1)
FIG. 1 shows an operation principle of an audio signal processing device 1 (audio signal preprocessing device 10) according to an embodiment of the present invention.

  As shown in FIG. 1, the audio signal processing device 1 includes an audio signal preprocessing device 10, a microphone 12, an audio recognition unit 16, a first noise reference sensor 28, and a second noise reference sensor 30. Yes. In FIG. 1, the audio signal preprocessing device 10 is configured to reduce external noise from the audio signal received by the microphone 12. The audio signal processing pre-processing apparatus 10 according to the first embodiment of the present invention is applied to one microphone 12 for cost reasons. However, if necessary, the audio signal processing pre-processing apparatus 10 may be applied to a more expensive network including a plurality of microphones. May be applied. The output 14 of the audio signal preprocessing device 10 is input to the audio recognition unit 16. The audio signal pre-processing device 10 is particularly configured to reduce some external noise generated in the vehicle. The output of the voice recognition unit 16 is used, for example, to generate an input signal to the in-vehicle electric device. In addition or alternatively, the voice signal pre-processing device 10 outputs filtered voice signals regardless of whether voice recognition is performed, and these voice signals are used in a vehicle communication system such as a mobile phone. The mobile phone includes a hands-free phone.

  The audio signal preprocessing device 10 is executed by a dedicated hardware circuit, hardware that can be constructed, a filtering algorithm executed by a processor, or a combination of any of these. Further, the audio signal pre-processing device 10 may be executed in an integrated circuit such as an application specific IC (ASIC), and may operate together with the voice recognition unit 16 in the same integrated circuit.

  In general, the acoustic signal x (n) received by the microphone 12 has at least one of the following components.

Speech signal component s (n): a component included while the speaker is speaking, and a desired signal to which noise is not added, which should be output to the speech recognition unit 16.
Coherent noise component c (n): a component coherent with at least one first noise reference signal, for example, an output of a vehicle audio system (eg, radio, audio, video player).

  Non-coherent noise component nc (n): a component that is non-coherent with the noise reference signal in the vehicle, and has at least one of the following components.

  Non-coherent vehicle component ncv (n): a component that is non-coherent with the second vehicle noise reference signal but is correlated with the second vehicle noise reference signal in the power spectrum. For example, engine noise and tire noise.

  External relatively stable component (external stationary component) d (n): Strictly speaking, it is not related to the second vehicle noise reference signal, but it changes relatively slowly with time. For example, road noise, rain sound and wind sound.

Transient component t (n): vehicle horn, noise from other vehicles, transient noise outside the vehicle, etc. Therefore, nc (n) = ncv (n) + d (n) + t (n)
In general, the audio signal preprocessing device 10 includes a first filter 20 and a second filter 22, and the first and second filters 20 and 22 are related to each other or executed in order. One of the first and second filters 20 and 22 is a coherent filter 24 that reduces noise according to a coherent noise reduction algorithm CNRA (for example, a linear filter). The coherent filter 24 reduces the coherent noise component c (n). The other of the first and second filters 20 and 22 is a non-coherent filter 26 that reduces noise according to a non-coherent algorithm (eg, non-linear filter, non-linear noise reduction algorithm NLNR). The non-coherent filter 26 reduces the non-coherent noise component nc (n) (that is, the non-coherent vehicle component ncv (n) or, optionally, the external stationary component d (n) may be added).

  In Embodiment 1 of the present invention, since the non-coherent filter 26 is executed after the coherent filter 24, the filtering performance is improved. Further, the non-coherent filter 26 is prevented from being detuned by a noise component that is more efficiently reduced by the coherent filter 24.

  By executing the coherent filter and the non-coherent filter in order, various external noise components can be reduced over a wider range (various types of noise) than in the past.

  The coherent filter 24 and the non-coherent filter 26 receive the first noise reference signal of the first noise reference sensor 28 and the second noise reference signal of the second noise reference sensor 30, respectively. These first and second noise reference signals are used to reduce coherent or non-coherent noise components. Since the first and second noise reference sensors 28 and 30 are non-acoustic reference sensors, interference with an audio signal received via the microphone 12 is avoided. The first noise reference sensor 28 for the coherent filter 24 is, for example, directly connected (for example, non-acoustic) to a speaker drive signal of a vehicle radio player and / or a music player, but may be electrically connected. The speaker drive signal is a monaural signal, a stereo signal, or a multidimensional signal (for example, surround sound). In the case of a stereo signal or a multidimensional signal, separate signals are applied as individual noise reference signals, so that a plurality of noise reference signals are input to the coherent filter 24. As described above, when a plurality of noise reference signals are input, the filtering performance of the coherent filter 24 increases, but the filtering may be complicated. On the other hand, when a plurality of or all received signals are downmixed to reduce the number of noise reference signals, the complexity of filtering can be reduced. For example, when a stereo signal is downmixed to a monaural signal, a single noise reference signal generated based on the original two stereo signal components is input to the filters 24 and 26. While signal downmixing will reduce the final performance characteristics of the coherent filter 24, in practice it has proven to be significantly effective in filtering performance. Such a downmix optimizes the cost to performance ratio within the desired performance level range. The plurality of signals may be appropriately downmixed or weighted as appropriate.

  The second noise reference sensor 30 for the non-coherent filter 26 is a non-acoustic sensor, for example, an accelerometer or a vibration sensor mounted on a vehicle body or a vehicle floor. FIG. 2 schematically illustrates the arrangement of the first noise reference sensor 28, the second noise reference sensor 30 and the microphone 12. The microphone 12 is arranged near the driver of the passenger compartment and receives the spoken word of the speaker (in this case, the driver). The first noise reference source 28 is directly connected to the vehicle audio system. The second noise reference sensor 30 is generally attached to a vehicle body or a vehicle floor and receives a mechanical signal.

  FIG. 3 illustrates the detailed structure of the audio signal preprocessing apparatus 10. The audio signal preprocessing device 10 includes an input unit 32 that receives signals received from the microphone 12 and signals supplied from the first noise reference sensor 28 and the second noise reference sensor 30. The input unit 32 includes a digital unit that digitizes (binarizes) a received signal when the received signal is not in a digital format. The input unit 32 includes a frame unit that divides the generated digital signal into overlapping frames. Each frame time is approximately 10 ms per frame, for example. The input unit 32 further includes a band separation unit that separates the received signal into N (predetermined value) frequency bands. For example, the band separation unit is executed using a fast Fourier transform (FFT). Frequency bands are described on a logarithmic scale within the frequency domain. As a result, the signal quality over several octaves can be maintained uniformly, and the processing can be prevented from becoming complicated. Signals received from the microphone 12, the first noise reference sensor 28, and the second noise reference sensor 30 are digitized by the input unit 32, and output from the input unit 32 in the digital form of signals 12a, 28a, and 30a. These signals 12a, 28a and 30a are divided into frames in the time domain and also divided into N frequency bands in the frequency domain. The frequency band number N is set to about 11 in the frequency region of 250 Hz to 6 kHz.

  The audio signal preprocessing apparatus 10 further includes an output unit 36 that generates an output signal by synthesizing the filtered N subband signals. The output unit 36 includes an inverse FFT (Inverse Fourier Transform) that converts the N subband signal from the frequency domain to the time domain, and an overlap adder that generates a signal.

  Although both filters can be applied as the coherent filter 24 and the non-coherent filter 26, the coherent filter 24 is a linear filter that reduces noise coherent with the first noise reference signal, and the non-coherent filter 26 is non-coherent with the second noise reference signal. It is a non-linear filter that reduces noise that is correlated in the power spectrum even in the coherent manner. 4 and 5 illustrate an example of the coherent filter 24. Referring to FIG. 4 (and considering only the coherent external noise supplied from the vehicle speaker), the signal of the microphone 12 (hereinafter referred to as microphone signal) x (n) is shown as follows.

x (n) = s (n) + g _c (a (n))
Here, s (n) is a sampled audio component signal, which is a desired signal that does not include external noise. a (n) is a signal transmitted to the speaker and is a first noise reference signal from the first noise reference sensor 28. g _c is a transfer function from the speaker to the microphone 12 and is linear. The time series signal is divided into overlapping frames, and each signal is subjected to FFT (Fourier transform), and then X _i (k) is obtained by the following expression in which the frequency is binarized for each frame i. X _i (k) = FFT (x _i ), S _i (k) = FFT (s _i ), A _i (k) = FFT (a _i ) and G ⁱ _c (k) = FFT (g ⁱ _c )
With such a notation, the coherent filter 24 is shown in FIG. Here, k is a variable indicating the number of bins corresponding to the frequency, and i indicating the number of frames is omitted for simplification.

X (k) indicates the microphone signal in the current frame,
S (k) indicates the audio signal without noise in the current frame,
G _c (k) represents the transfer function between the speaker and microphone in the current frame,
A (k) indicates the audio reference signal in the current frame,
H (k) represents the estimated value of the transfer function G _c (k) by the improved Wiener algorithm in the current frame, and Equation 1 is the output value of the coherent filter 24 in the current frame, and the speech signal with enhanced speech Indicates.

  The function H (k) is based on a noise reduction method called a Wiener filter in the frequency domain. The acoustic transmission from the plurality of speakers to the microphone 12 is considered to be linear. The algorithm for estimating the function H (k) is an iterative process rather than an adaptive process. When using iterative processing, the filter update depends only on the input values A (k) and X (k) and is independent of the output value (Equation 1). Therefore, the stability of filter update is improved.

  For example, the filter coefficient H is repeatedly estimated by the filter coefficient calculation unit 45 using the following algorithm.

H (k) = γ _XA (k) γ ^-1 _AA (k)
Where γ _XA (k) indicates the estimated value of the cross spectrum between the microphone signal and the audio reference signal, γ _AA (k) indicates the estimated value of the audio reference signal spectrum, and γ _XA (k) and γ _AA (k) is repeatedly estimated in the current frame and depends on the respective values in the previous frame, the instantaneous spectrum value and the cross spectrum value in the current frame.

  The coherent filter 24 further comprises or is associated with a first switching unit 34 that determines whether the vehicle music player / radio is currently ON. The switching unit 34 controls so that only the coherent filter 24 operates when the vehicle music player / radio is in the ON state. For example, the switching unit 34 directly receives the “ON / OFF” display signal supplied from the vehicle music player / radio, or determines whether the value of the signal 28a of the first noise reference source 28 is greater than a certain threshold value. You may have a threshold switch to do. The switching unit 34 performs single on / off control for all bands of the coherent filter 24 or performs individual control for each band.

(Embodiment 2)
In the second embodiment, an example of the non-coherent filter 26 will be described with reference to FIGS. This example is particularly suitable for use in the audio signal pre-processing device 10 of the first embodiment, but the second embodiment is not limited to this, and the non-coherent filter 26 is related to noise reduced by spectral power. The present invention can be applied to a device (particularly an in-vehicle device) that reduces noise based on a noise reference signal.

  The feature of the second embodiment is that filtering is performed without using a voice interval detector (VAD) module. The VAD detector is generally required for the conventional spectral subtraction filter, but has a problem of stability.

  Before the detailed description of the second embodiment, the principle of spectral subtraction for one channel having a VAD module will be described for better understanding. The principle of 1-channel spectral subtraction is as follows.

  As shown in FIG. 10, x (n) represents a sample signal from a microphone containing sample sound s (n) and noise b (n).

The output value (Equation 2) is an enhanced audio signal.

When the same symbol as described above is used for each frame in the same frequency region, the output value (Equation 2) is obtained as Equation 3.

Here, G (k) (Equation 5) is a gain function, and Equation 4 is an estimated noise value during speechlessness.

  h (.) is a function based on a change in spectrum. These are disclosed in (S.V.Vaseghi, "Advanced Digital Signal Processing and Noise Reduction", published by John Wiley & Sons Ltd, 2000).

  While one-channel spectral subtraction is effective, it relies heavily on VAD performance to distinguish between noisy speech frames and noisy frames.

In the second embodiment shown in FIG. 6, x (n) represents a sample signal transmitted from the microphone 12, but the coherent component already reduced by the coherent filter 24 of the embodiment is included in this sample signal. Absent. In the embodiment shown in FIG. 6, the illustration and description of the coherent filter 24 are omitted, but these two filters 24 and 26 may be executed continuously as in the first embodiment. x (n) contains a sample speech component s (n) and a non-coherent noise component nc (n). nc (n) is composed of a non-coherent vehicle component ncv (n), an external stationary component d (n), and a transient component t (n). The non-coherent vehicle component ncv (n) received by the microphone 12 is related to the signal r (n) received by the second noise reference sensor 30 by the nonlinear function f _NC : ncv (n) = f _NC (r (n)). It has been. After dividing the time series signal into frames that overlap each other little by little, and applying Fourier transform (FFT) to each signal, X _i (k) is obtained for each frame i and frequency bin k.
X _i (k) = FFT (x _i ), S _i (k) = FFT (s _i ), NCV _i (k) = FFT (ncv _i ), D _i (k) = FFT (d _i ), T _i (k) = FFT (t _i )
Nonlinear filter 26 is shown in FIG. 7 using these notations. Here, k is a variable representing a bin corresponding to a frequency, and i indicating a frame number is omitted for simplification.

X (k) = S (k) + NCV (k) + D (k) + T (k)
Assuming that much of the noise received by the microphone 12 is also received by the vibration sensor, the vibration sensor is used as a noise reference sensor and reduces the noise contained in the signal that is the output of the linear filter by a technique based on spectral subtraction. . Here, the noise component and the vibration reference signal component output from the linear filter are incoherent, but the point is that they are related to each other in the power spectrum.

The non-coherent filter 26 includes a gain unit 40 that performs spectral subtraction by applying the spectral subtraction function G _NC to the microphone signal X. The spectral subtraction function G _NC is a nonlinear function expressed as G _NC (k) = G _NC [R (k), X (k), ref_calib]. Since the non-coherent filter 26 is a non-linear filter, it is executed to cancel noise by spectral subtraction. G _NC is similar to the gain used in 1-channel spectral subtraction. The estimated speech signal (Equation 6) is calculated as shown in Equation 7.

The gain unit 40 receives the microphone signal X (k) and a new noise reference signal 42ref_calib ^* R (k) that is the noise reference signal R (k) amplified by the calibration gain ref_calib. The calibration gain ref_calib is calculated by the estimation unit 44, for example. The calibration gain “ref_calib” is an estimated value of the square value of the transfer function between the reference signal sensor and the microphone, or a value related thereto. The calibration gain ref_calib is a single value or signal, or has a spectral component. The non-coherent filter 26 includes a multiplication unit 46 that multiplies the signal R (k) of the second noise reference source by the calibration gain ref_calib. The estimation of the calibration gain “ref_calib” is continuously performed by the update algorithm in accordance with the following principle.

(a) The transfer function between the noise reference signal and the microphone changes in a relatively slow time (for example, on the order of seconds). During normal operation of the vehicle, the spectral fluctuation range of the microphone signal is approximately proportional to the spectral fluctuation range of the signal of the second noise reference sensor 30. Moreover, even if the microphone signal and the signal of the second noise reference sensor 30 each change rapidly, the spectral fluctuation ranges are still proportional to each other. This is because the non-coherent vehicle component NCV (k) and the noise reference signal R (k) are related to each other in the power spectrum.

(b) The external relatively stable component D (k) is assumed to vary relatively slowly and with vehicle speed. The external relatively stable component D (k) is adjusted by the magnification λ of the calibration gain ref_calib.

(c) Although there are several methods for estimating the calibration gain ref_calib, it may be estimated by the following equation. ref_calib = λE _x / E _r
Here, E _x denotes the estimated value of the instantaneous power of the microphone signal. For example, _Ex is obtained as shown in Equation 8.

Here, the frame L usually lasts in seconds, for example, 0.5 seconds. _Er represents an estimated value of the instantaneous power of the second noise reference signal. For example, E _r is obtained as shown in Equation 9 below.

  Here, the frame L usually lasts in seconds, for example, 0.5 seconds. λ (1) is an element used not to overestimate the ratio (contribution) of the non-coherent vehicle component NCV (k) and to take into account the relatively stable external component D (k) . λ generally varies in the range of about 0.7 to about 1, with the rate of variation depending on the vehicle speed.

  The magnitude of the power spectrum is generally estimated about every 0.5 to 1 second, and the calibration gain ref_calib is generally estimated about every 1 to about 3 seconds.

(d) Since the transfer function (between the noise reference sensor and the microphone) and | D (k) | ² change relatively slowly, there is no difference between the magnitudes of the squares of both signals X (k) and R (k). If a balanced difference occurs, this indicates that there is external confusion in the audio signal S (k) or transient noise T (k). By applying the change threshold or the change rate threshold to the calibration gain ref_calib obtained by the estimation unit 44, the calibration gain ref_calib can be prevented from being distorted by such external noise. For example, when the value of the calibration gain ref_calib changes more than about 20% from the previous value, the change is not reflected and the previous value of the calibration gain ref_calib is used instead.

  According to this technique, in the second embodiment, it is not necessary to use VAD, and it is not necessary to use a transient noise detector. By using the change threshold instead of these, the calibration gain ref_calib can be controlled globally, and disturbance due to the audio signal S (k) or the transient noise T (k) can be avoided. Since VAD has a problem for effective implementation in conventional processing circuits, it is technically very significant that there is no need for VAD.

(e) The change threshold may be selected as follows.

(A) The change threshold is selected so that the calibration gain ref_calib is sufficiently large to follow a normal change in the transfer function value between the vibration sensor and the microphone.

(B) The change threshold is selected to a sufficiently small value so that the calibration gain ref_calib is not disturbed by a signal component having a change rate that cannot change because the change rate of the transfer function is too fast. . As described above, a change threshold of about 20% is effective.

  As described above, the calibration gain is limited by the limiting unit 44 so that the change rate of the calibration gain does not change beyond a predetermined threshold.

  As is clear from FIG. 8, the non-coherent filter 26 is divided and set to have a plurality of frequency bands as in the first embodiment. The non-coherent filter 26 has an input unit 50 that divides an input value into frames for each time and divides an input signal x (n) and an input signal r (n) from a non-coherent noise reference sensor into frequency bands. The input unit 50 includes an FFT unit that divides these signals into frequency bands. The non-coherent filter 26 includes an output unit 52 that synthesizes signals from the plurality of sub-filter units 26 ′. The output unit 52 has an inverse FFT unit. When the non-coherent filter 26 of the second embodiment is included in the first embodiment, the input unit 50 and the output unit 52 may be executed instead by the input unit 32 and the output unit 36 of the first embodiment. It may not be executed again in the coherent filter 26.

  FIG. 9 shows an arrangement example of the microphone 12 and the non-acoustic non-coherent sensor 30 in the vehicle when the second embodiment is implemented independently. The microphone 12 is installed near the driver and acoustically receives the driver's voice. A non-acoustic non-coherent noise reference signal sensor 30 is typically provided on the vehicle body or vehicle floor and receives mechanical vibrations. The output signal of the non-coherent filter 26 is transmitted to a speech recognition circuit (not shown) as in the first embodiment.

The present invention provides a speech preprocessing method for filtering an electrical speech signal including speech and reducing a noise component, wherein the electrical speech signal is subjected to either a coherent filtering or a non-coherent filtering by a first filter. Process,
A method may be provided that includes a second filtering step of filtering the output based on the first filter with the other filter different from the first filter by the second filter.

  The present invention may also be an audio signal preprocessing program that causes a computer to execute the above method.

FIG. 1 is a schematic view schematically showing the principle of an audio signal preprocessing device according to the first embodiment. FIG. 2 is a schematic diagram schematically showing the arrangement of the microphone and the non-acoustic noise reference sensor in the vehicle. FIG. 3 is a block diagram schematically showing details of the audio signal preprocessing device. FIG. 4 is a schematic diagram schematically showing the principle of coherent filtering. FIG. 5 is a schematic view schematically showing details of the structure of the coherent filter. FIG. 6 is a schematic diagram schematically illustrating the principle of non-coherent filtering according to the second embodiment. FIG. 7 is a block diagram schematically showing a non-coherent filter. FIG. 8 is a block diagram schematically showing details of the structure of the non-coherent filter. FIG. 9 is a diagram schematically illustrating the arrangement of the microphone and the non-acoustic noise reference sensor when the second embodiment is used independently of the first embodiment. FIG. 10 is a diagram schematically showing the principle of 1-channel spectral subtraction.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Audio | voice signal processing apparatus 10 Audio | voice signal pre-processing apparatus 12 Microphone 16 Voice recognition part 20 1st filter 22 2nd filter 28 Non-acoustic coherent noise sensor 30 Non-acoustic non-coherent noise sensor

Claims (21)

An audio signal pre-processing device that includes a noise reducing first filter and a noise reducing second filter and that processes an electric audio signal including audio to reduce a noise component, wherein the first filter or the second filter Either one is a coherent filter that reduces the noise component by coherent filtering, and the other is a non-coherent filter that reduces the noise component by non-coherent filtering, and the second filter filters the output of the first filter. An audio signal preprocessing device. The audio signal preprocessing device according to claim 1, wherein the first filter is the coherent filter, and the second filter is the non-coherent filter. The coherent filter is configured to receive a first noise reference signal that is coherent with a noise component included in the electrical audio signal from a non-acoustic coherent noise sensor. Audio signal pre-processing device. The non-coherent filter is configured to receive from a non-acoustic non-coherent noise sensor a second noise reference signal that is non-coherent with a noise component included in the electrical audio signal and is related in spectral power. The audio signal preprocessing device according to any one of claims 1 to 3, wherein The audio signal processing apparatus according to claim 3 or 4, wherein the non-acoustic coherent noise sensor is non-acoustically connected to an audio system of a vehicle. 6. The audio signal processing apparatus according to claim 4, wherein the non-acoustic non-coherent noise sensor is configured to sense vehicle vibration. The audio signal preprocessing apparatus according to any one of claims 1 to 6, wherein the coherent filter includes a linear filter. The audio signal preprocessing apparatus according to claim 1, wherein the coherent filter includes a Wiener filter. The sound according to any one of claims 3 to 8, wherein the coherent filter includes a filter coefficient calculation unit that repeatedly calculates a filter coefficient based on the electric sound signal and the first noise reference signal. Signal pre-processing device. The audio signal preprocessing device according to any one of claims 1 to 9, wherein the non-coherent filter includes a non-linear filter. The audio signal preprocessing device according to any one of claims 1 to 10, wherein the non-coherent filter includes a spectrum-gain filter. 12. The non-coherent filter further includes a filter coefficient calculation unit that calculates a filter coefficient based on the electric audio signal and the second noise reference signal input to the non-coherent filter. The audio signal preprocessing device according to any one of the above. Each of the first filter and the second filter has a plurality of sub-filters that filter the electrical audio signal in a plurality of corresponding frequency sub-bands. The audio signal pre-processing device described. The audio signal according to claim 13, further comprising: an input unit that divides the electric audio signal into the frequency subband; and an output unit that synthesizes the electric audio signal from the frequency subband after filtering. Pre-processing device. The first filter and the second filter process the electrical audio signal as a series of frames, and each frame processes the signal as a continuous frame of frames representing a signal in a window for a predetermined period. The audio signal preprocessing device according to any one of claims 1 to 14. 16. The audio signal preprocessing according to claim 15, further comprising: an input unit that divides the electric audio signal into the series of frames; and an output unit that synthesizes the electric audio signal from the series of frames. apparatus. The audio signal pre-processing device according to any one of claims 1 to 16,
A microphone that collects an acoustic signal and supplies the electrical audio signal as an input to the audio signal preprocessing device;
The non-acoustic coherent noise sensor that generates the first noise reference signal coherent with a first noise source that can be collected by the microphone, and supplies the first noise reference signal to the coherent filter;
Generating a second noise reference signal that is non-coherent with a second noise source that can be collected by the microphone and is related in spectrum power to the second noise source; An audio signal processing apparatus comprising: the non-acoustic non-coherent noise sensor supplied to a non-coherent filter. The audio signal processing apparatus according to claim 17, wherein the non-acoustic non-coherent noise sensor is a vibration sensor or an accelerometer. In a speech preprocessing method for filtering an electrical speech signal including speech and reducing a noise component,
A first filtering step of filtering the electrical audio signal with a first filter, either coherent filtering or non-coherent filtering;
And a second filtering step of filtering the output based on the first filter by the second filter, which is the other filtering different from the first filter. The audio signal preprocessing method according to claim 19, wherein the first filter is the coherent filter, and the second filter is the non-coherent filter. An audio signal preprocessing program for causing a computer to execute the method according to claim 19 or 20.

Images