JP6613728B2 - Noise suppression device, program and method - Google Patents

Description

  The present invention relates to a noise suppression device, a program, and a method, and can be applied to noise suppression of input signals of a plurality of channels input from a plurality of microphones arranged at different positions in a common acoustic space, for example.

  Since noise exists everywhere in the natural environment, generally, when recording voice in the real world, noise from various sources is mixed in the observed signal. Those noises reduce the intelligibility of speech even when a person listens. In addition, when a signal including noise is input to a speech processing device such as a speech recognition device, a sound image localization device, or a sound source separation device, the performance of the speech processing (for example, speech recognition rate, accuracy of sound image localization, sound quality of sound source separation sound) Reduce. For this reason, there is a great demand for technology that enhances the speech component (target sound component) by suppressing the noise component (non-target sound component) mixed in the input signal, and so far various noise suppression methods (speech enhancement methods) Has also been developed.

  Here, terms used in the present specification are defined as follows. Hereinafter, a signal input to the noise suppression device is referred to as an “input signal”, and a frequency analysis (converted into the frequency domain) of the input signal is referred to as an “input spectrum”. In the following, the absolute value of each element of the input spectrum is referred to as “input amplitude spectrum”, and the power of each element of the input spectrum (for example, the square of the absolute value) is calculated as “input power spectrum”. ". Furthermore, in the following, the spectrum in which noise is suppressed by the noise suppression device is referred to as “post-suppression spectrum”, and the absolute value of each element of the post-suppression spectrum is referred to as “post-suppression amplitude spectrum”. The power of each element is calculated as a “suppressed power spectrum”. Furthermore, hereinafter, a signal waveform restored from the post-suppression spectrum is referred to as a “post-suppression signal”. Hereinafter, the spectrum of the noise component included in the input signal is referred to as “noise amplitude spectrum”, and the power of each element of the noise amplitude spectrum is calculated as “noise power spectrum”.

  By the way, conventionally, there are two problems in the method of performing noise suppression for each channel for multi-channel signal processing using input signals of a plurality of channels such as sound image localization and sound source separation. The first problem is that when the independently calculated suppression gain is applied to each channel, the power balance between the channels is lost and the performance of multichannel signal processing is degraded. The second problem is that the amount of processing increases as the number of channels increases because the suppression gain is calculated for each channel.

  As a conventional technique for dealing with the above problems, there is a technique described in Patent Document 1. In the technique described in Patent Document 1, one representative signal is generated from input signals of a plurality of channels, and the suppression gain calculated based on the representative signal is commonly applied to each channel. I'm trying to solve the problem of increased throughput. In the apparatus described in Patent Document 1, the representative signal is generated by averaging all channels of the input power spectrum, weighted addition, selection of maximum power, and selection of minimum power. In the apparatus described in Patent Document 1, the suppression gain is calculated based on the representative signal obtained in this way. In the apparatus described in Patent Document 1, since the suppression gain is calculated only once, the amount of processing can be reduced compared to calculating the suppression gain for each channel. In the apparatus described in Patent Document 1, the obtained suppression gain is commonly applied to each channel, so that the power balance between the channels is maintained.

Japanese Patent No. 5435204

  However, in the apparatus described in Patent Document 1, the processing amount for generating the representative signal requires an O (n) processing amount order (processing in proportion to the number of channels) according to the number of channels. . For example, M is the number of channels and K is the number of elements of the input power spectrum. At this time, when a representative signal is obtained by weighted addition using the device described in Patent Document 1, M * K multiplications are required. Moreover, in the apparatus described in Patent Document 1, when obtaining a representative signal by selecting the maximum power, (M−1) * K comparisons are necessary.

  Therefore, a noise suppression device, a program, and a method that can efficiently perform processing when performing noise suppression processing on a plurality of channels of input signals are desired.

According to a first aspect of the present invention, there is provided a noise suppression apparatus that suppresses noise from an input spectrum corresponding to each of a plurality of channel input signals. (1) One channel from the plurality of channels according to a predetermined method every predetermined unit time. (2) a suppression gain calculation means for calculating a suppression gain for suppressing a noise component included in the input spectrum using the input spectrum of the selected channel, and (3) the suppression Noise suppression means for suppressing a noise component for each of the input spectra of the plurality of channels using a gain is provided.

A noise suppression program according to a second aspect of the present invention provides a computer mounted on a noise suppression apparatus that suppresses noise from an input spectrum corresponding to each of a plurality of channels of input signals . suppression gain calculating the channel selection means for selecting one channel from the plurality of channels by method a suppression gain for suppressing a noise component contained in the input spectrum by using the input spectrum of the channel selected (2) And (3) functioning as noise suppression means for suppressing noise components for each of the input spectra of the plurality of channels using the suppression gain.

According to a third aspect of the present invention, in a noise suppression method for suppressing noise from an input spectrum corresponding to each of a plurality of channel input signals, (1) a channel selection unit, a suppression gain calculation unit, and a noise suppression unit are provided. ) The channel selection means selects one channel from the plurality of channels by a predetermined method every predetermined unit time. (3) The suppression gain calculation means uses the input spectrum of the selected channel to Calculating a suppression gain for suppressing a noise component included in the input spectrum, and (4) the noise suppression means suppresses the noise component for each of the input spectra of the plurality of channels using the suppression gain. Features.

  ADVANTAGE OF THE INVENTION According to this invention, the noise suppression apparatus which performs efficiently the process at the time of performing a noise suppression process about the input signal of multiple channels can be provided.

It is the block diagram shown about the functional structure of the noise suppression apparatus which concerns on 1st Embodiment. It is the block diagram shown about the internal structure of the suppression means classified by band which concerns on 1st Embodiment. It is the block diagram shown about the internal structure of the channel selection means which concerns on 1st Embodiment. It is the block diagram shown about the internal structure of the channel determination means which concerns on 1st Embodiment. It is the block diagram shown about the internal structure of the suppression gain calculation means which concerns on 1st Embodiment. It is the block diagram shown about the internal structure of the channel selection means which concerns on 2nd Embodiment. It is the block diagram shown about the internal structure of the channel determination means which concerns on 2nd Embodiment. It is the block diagram shown about the internal structure of the channel selection means which concerns on 3rd Embodiment. It is the block diagram shown about the internal structure of the channel determination means which concerns on 3rd Embodiment. It is the figure shown about the data structure of the threshold value set which concerns on 3rd Embodiment. It is the flowchart shown about operation | movement of the channel determination means which concerns on 3rd Embodiment.

(A) First Embodiment Hereinafter, a first embodiment of a noise suppression device, a program, and a method according to the present invention will be described in detail with reference to the drawings.
(A-1) Configuration of First Embodiment FIG. 1 is a block diagram showing a functional configuration of a noise suppression device 100 of this embodiment. In FIG. 1, reference numerals in parentheses are used only in second and third embodiments described later.

  The noise suppression apparatus 100 receives input of input signals is1 to isM (time domain signals) of M channels. In the following, it is assumed that a channel number is assigned to each of the input signals is1 to isM as a channel identifier for identifying a channel. For example, the channel number of the input signal is1 is 1, the channel number of the input signal is2 is 2,..., And the channel number of the input signal isM is M. The input signals is1 to isM are audio signals (acoustic signals) captured by different microphones (not shown), for example. Assume that microphones (not shown) related to the input signals is1 to isM are arranged in a limited range such as in the same room.

  Then, the noise suppression device 100 outputs post-suppression signals os1 to osM (time domain signals) corresponding to the input signals is1 to is as outputs after the noise suppression processing. In this embodiment, the noise suppression apparatus 100 is described as receiving a time-domain signal and outputting a time-domain signal. However, the format of the input / output signal is not limited. For example, Also, a configuration may be adopted in which signal input in the frequency domain (spectrum) or signal output in the frequency domain is performed.

  The noise suppression apparatus 100 outputs frequency analysis means 101 (101-1 to 101-M), band-specific suppression means 102, and post-suppression signals os1 to osM corresponding to the input signals is1 to isM, respectively. And restoring means 103 (103-1 to 103-M).

  The noise suppression device 100 may be constructed, for example, by installing a computer program (a noise suppression program according to the embodiment) in a computer having a processor and a memory (program implementation configuration). The noise suppression apparatus 100 may be realized partially or entirely using hardware (for example, a dedicated semiconductor chip or an electric circuit).

  The frequency analysis units 101-1 to 101-M perform processing for converting the input signals is1 to isM from time domain signals to frequency domain signals, respectively. Hereinafter, signals obtained by converting the input signals is1 to isM into the frequency domain (signals output from the frequency analysis means 101-1 to 101-M) are represented as input spectra IS1 to ISM, respectively. For example, the frequency analysis unit 101-1 converts the frequency of the input signal is 1 to generate and output the input spectrum IS 1. Although the specific method of the conversion process performed by each frequency analysis unit 101 is not limited, for example, fast Fourier transform (FFT), wavelet transform, filter puncture, or the like may be applied. In this embodiment, each frequency analysis unit 101 will be described as performing a conversion process using FFT. In the following description, each input spectrum IS1 to ISM is assumed to be data represented by complex numbers.

  The band-by-band suppression unit 102 performs a process of suppressing noise components for each band for the input spectra IS1 to ISM converted to the frequency domain, and the post-suppression spectrums OS1 to OSM are obtained as noise component suppressed signals (frequency domain signals). Is generated. The band suppression unit 102 supplies the suppressed spectra OS1 to OSM to the waveform restoration units 103-1 to 103-M, respectively. For example, the band-by-band suppression unit 102 performs a process of noise suppression on the input spectrum IS1 to generate a post-suppression spectrum OS1 and supplies it to the waveform restoration unit 103-1.

  The waveform restoration means 103-1 to 103 -M restore the post-suppression spectra OS <b> 1 to OSM supplied from the band-by-band suppression means 102, respectively, to time domain signals and output them as post-suppression signals os <b> 1 to osM. For example, the waveform restoration unit 103-1 converts the post-suppression spectrum OS <b> 1 into the time domain and outputs the post-suppression signal os <b> 1. Each waveform restoration unit 103 converts each suppressed spectrum into a signal in the time domain by a method paired with the frequency analysis units 101-1 to 101 -M. For example, when frequency conversion using FFT is performed by the frequency analysis units 101-1 to 101-M, the waveform restoration units 103-1 to 103-M use inverse fast Fourier transform (Inverse FFT: IFFT). Thus, each suppressed spectrum is restored to a time domain signal.

  Next, the internal configuration of the band-by-band suppression unit 102 will be described with reference to FIG.

  FIG. 2 is a block diagram showing the internal configuration of the band-by-band suppression unit 102.

  As shown in FIG. 2, the band-by-band suppression unit 102 includes a channel selection unit 201, a suppression gain calculation unit 202, and a multiplication unit 203 (203-1 to 203-M).

  All of the input spectra IS1 to ISM supplied to the band-by-band suppression unit 102 are supplied to the channel selection unit 201 and also supplied to the multiplication units 203-1 to 203-M. For example, the input spectrum IS1 is supplied to the multiplication unit 203-1.

  The channel selection means 201 performs a process of selecting and outputting any one of the input spectra IS1 to ISM by a predetermined method. Hereinafter, the input spectrum selected by the channel selection unit 201 (selected one of the input spectra IS1 to ISM) is referred to as a selected input spectrum. The channel selection unit 201 supplies the selected input spectrum to the suppression gain calculation unit 202.

  The channel selection unit 201 performs processing for switching the channel number to be the selected input spectrum in a predetermined order. For example, the channel selection unit 201 switches the channel number to be the selected input spectrum every unit time (for example, the processing unit time of the suppression gain calculation unit 202). At this time, the order in which the channel selection unit 201 switches the channel numbers is not limited. In this embodiment, the channel selection means 201 shall select a channel number cyclically in order of a channel number (select a channel number used as a selection input spectrum). In other words, it is assumed that the channel selection unit 201 selects a channel number as a selection input spectrum in round-robin order in the order of channel numbers. That is, it is assumed that the channel selection unit 201 of this embodiment selects channel numbers to be selected input spectra in the order of “1, 2,..., M, 1, 2,.

  Note that the algorithm of the order in which the channel selection unit 201 selects the channel number to be the selected input spectrum is not limited. For example, in addition to the example of selecting a channel number as a cyclically selected input spectrum in the order of channel numbers, the order may be set in advance as in the second embodiment described later, or as in the third embodiment described later. May be selected in an order based on pseudo-random numbers (random order).

  The suppression gain calculation unit 202 estimates and acquires the noise component spectrum based on the selected input spectrum. Hereinafter, the noise spectrum estimated by the suppression gain calculation unit 202 is also referred to as “suppression gain”. Details (detailed configuration) of the processing in which the suppression gain calculation unit 202 calculates the suppression gain will be described later. The suppression gain calculation unit 202 supplies the calculated suppression gain to each multiplication unit 203-1 to 203-M.

  Multiplication means 203-1 to 203-M perform a process of multiplying the input spectra IS1 to ISM by a common suppression gain (a suppression gain supplied from the suppression gain calculation means 202), respectively. Thereby, the multiplication means 203-1 to 203-M suppress the noise components included in the input spectra IS1 to ISM, respectively, and obtain post-suppression spectra OS1 to OSM.

  Then, the multiplication means 203-1 to 203-M output the suppressed spectra OS1 to OSM, respectively. For example, the multiplying unit 203-1 performs a process of multiplying the input spectrum IS1 by a suppression gain and acquiring and outputting the suppressed spectrum OS1. Note that the suppressed spectra OS1 to OSM are supplied to the waveform restoration units 103-1 to 103 -M, respectively. That is, the multiplication means 203-1 to 203-M function as noise suppression means for performing noise suppression using a suppression gain common to each of the input spectra IS1 to ISM.

  Next, the internal configuration of the channel selection unit 201 will be described with reference to FIG.

  FIG. 3 is a block diagram showing the internal configuration of the channel selection means 201.

  As shown in FIG. 3, the channel selection unit 201 includes a frame counting unit 301, a channel determination unit 302, a channel storage unit 303, and an input spectrum selection unit 304.

  The frame counting unit 301 increments the processing frame number by one for every predetermined unit time (for example, the processing unit time of the suppression gain calculation unit 202), and gives the obtained frame number to the channel determination unit 302.

  The channel determination unit 302 selects the next channel number (next as the selected input spectrum) based on the given frame number and the channel number (channel number one unit time before) supplied from the channel storage unit 303. Channel number) to be determined. Then, the channel determination unit 302 gives the determined channel number to the input spectrum selection unit 304 as the selected channel number. The detailed configuration of the channel determination unit 302 will be described later.

  The channel storage means 303 stores the channel number given from the channel determination means 302 and gives the stored channel number to the channel determination means 302 after one unit time.

  The input spectrum selection unit 304 selects an input spectrum corresponding to the channel number given from the channel determination unit 302 as a selected input spectrum. For example, when the channel number given from the channel determination means 302 is m, the selected input spectrum handles the input vector of the channel m (input spectrum ISm) as the selected input spectrum.

  Next, the internal configuration of the channel determination unit 302 will be described with reference to FIG.

  FIG. 4 is a block diagram showing the internal configuration of the channel determination unit 302.

  As shown in FIG. 4, the channel determination unit 302 includes a channel counting unit 401, a constant 1 value supply unit 402, and an initial channel switch unit 403.

  The frame number given to the channel determining means 302 is given to the initial channel switching means 403. Further, the channel number one unit time before given to the channel determining means 302 is given to the channel counting means 401.

  The channel counting means 401 updates the given channel number one unit time ago, and gives the obtained provisional selection channel number to the initial channel switching means 403. The channel number is updated in the channel counting means 401 as follows. When the channel number before the unit time is 1 or more and (M−1) or less, the channel counting means 401 sets a value obtained by incrementing the channel number before the unit time by 1 as the provisional selection channel number. Then, when the channel number before the unit time is M, the channel counting unit 401 sets the temporary selection channel number to 1 as an initial value.

  The constant 1 value supply means 402 always gives the value “1” to the initial channel switch means 403. The value “1” is the first value (initial value) of the channel number. In the first embodiment, the initial value is the value “1”, but any value may be applied as the initial value as long as it is an integer of 1 or more and M or less representing the channel number.

  If the frame number is the initial frame number, the initial channel switch means 403 uses the value “1” given from the constant 1 value supply means 402 as the selected channel number, otherwise the provisional selection channel given from the channel counting means. The number is the selected channel number. The initial frame number varies depending on the implementation, and may be set to 0 or 1, for example. In order to stabilize the initial operation, the initial channel switch means 403 selects the value “1” given from the constant 1 value supply means 402 when the frame number is a small integer value (for example, 4) or less as the selected channel. It is good also as a number.

  Next, the internal configuration of the suppression gain calculation unit 202 will be described with reference to FIG.

  FIG. 5 is a block diagram showing the internal configuration of the suppression gain calculation unit 202.

  In this embodiment, the suppression gain calculation means 202 uses the noise power spectrum estimated based on the speech section determined based on the posterior SNR, and further uses the Wiener filter (WF) based on the a priori SNR estimated by the Decision-Directed method. ) To calculate (estimate) the suppression gain. A specific calculation method for calculating (estimating) the suppression gain in the suppression gain calculation unit 202 is not limited.

  For example, the suppression gain calculation means 202 uses weighted noise estimation (Reference 1 “IEICE Trans. On Fundam., Vol. E85-A, No. 7, PP. 1710-1718, Jul. 2002 ”) and a method based on MAP estimation (see Reference 2“ Technical Report of the Institute of Electronics, Information and Communication Engineers, Vol. 113, No. 503, EA 2013-121, PP. 7-12, March 2014 ”), etc. The used suppression gain calculation (estimation) processing may be performed, and the suppression gain calculation means 202 may be a spectral subtraction method (reference document 3 “IEEE Trans., Vol. ASSP-27, No. 2, pp. 113-120, Apr. 1979 ") and the MMSESTSA method (reference document 4" IEEE Trans., Vol. ASSP-32, No. 6, PP. 1109-1121, Dec. 1984 ") may be used to perform suppression gain calculation (estimation) processing.

  As shown in FIG. 5, the suppression gain calculation unit 202 includes a power calculation unit 501, a noise smoothing unit 502, a noise storage unit 503, a post-SNR calculation unit 504, a post-SNR storage unit 505, a threshold determination unit 506, and a threshold supply unit. 507, an a priori SNR estimation unit 508, a suppression gain determination unit 509, and a suppression gain storage unit 510.

  The suppression gain calculation unit 202 gives the given selection input spectrum to the power calculation unit 501. The power calculation unit 501 calculates the square of the absolute value of each element of the given selected input spectrum for each element, and gives the obtained input power spectrum to the noise smoothing unit 502 and the post SNR calculation unit.

Based on the noise power spectrum N ′ (k) before unit time given from the noise storage means 503 and the speech section judgment result V given from the threshold judgment means 506, the noise smoothing means 502 performs noise according to the following equation (1). The power is estimated, and the obtained noise power spectrum N (k) is given to the noise storage means 503 and the posterior SNR calculation means 504. In the following equation (1), k (k = 0,..., K−1) is an index of an element (frequency band), τ is a time constant, and X (k) is an input power spectrum. In the following equation (1), the time constant τ is preferably about 0.8.

  The noise storage means 503 stores the given noise power spectrum, and gives the stored noise power spectrum to the noise smoothing means 502 after one unit time.

  The posterior SNR calculation means 504 divides the input power spectrum given for each element by the given noise power spectrum, and gives the obtained posterior SNR to the posterior SNR storage means 505 and the prior SNR estimation means 508.

  The posterior SNR storage unit 505 stores the given posterior SNR, and provides the stored posterior SNR to the threshold determination unit 506 and the prior SNR estimation unit 508 after one unit time.

  The threshold determination unit 506 determines the voice segment determination result V if the posterior SNR before the unit time given from the posterior SNR storage unit 505 is larger than the posterior SNR threshold given from the threshold supply unit 507 (details will be described later). True (true value) is set to V. Otherwise, False (false value) is set to V. Then, the threshold determination unit 506 gives the obtained speech segment determination result V to the noise smoothing unit 502.

  The threshold supply unit 507 gives a predetermined posterior SNR threshold to the threshold determination unit 506. The posterior SNR compared with the posterior SNR threshold indicates that the input power is greater than the noise power if it is greater than 1, and the input power is smaller than the noise power if it is less than 1, the input power varies finely per unit time. However, the noise power does not change that much. Therefore, in order for the noise smoothing means 502 to update the noise power even if the input power is slightly larger than the noise power, it is preferable to set the posterior SNR threshold to a value of about 2.0.

  The a priori SNR estimation unit 508 calculates (estimates) the a priori SNR based on the given a posteriori SNR, the a posteriori SNR before unit time, and a suppression gain before unit time described later, and uses the obtained a priori SNR as a suppression gain determination unit. 509. Assuming that the posterior SNR is γ (k), the posterior SNR before unit time is γ ′ (k), the suppression gain before unit time is G ′ (k), and the prior SNR is ξ (k), the prior SNR is estimated as Decision It can be performed by the following equation (2) called the -Directed method. In the following formula (2), α is a forgetting factor, and is preferably set to a value of about 0.99.

The suppression gain determination unit 509 calculates the suppression gain G (k) by the following equation (3) based on the prior SNRξ (k), and gives the obtained suppression gain to the suppression gain storage unit 510. The suppression gain is output as an output of the suppression gain calculation unit 202.

  The suppression gain storage unit 510 stores the given suppression gain G (k), and provides the stored suppression gain to the prior SNR estimation unit 508 after one unit time.

(A-2) Operation of the First Embodiment Next, the operation of the noise suppression device 100 of the first embodiment having the above configuration (the noise suppression method of this embodiment) will be described with reference to FIGS. It explains using.

  The frequency analysis units 101-1 to 101-M perform processing for converting the input signals is1 to isM from time domain signals to frequency domain signals, respectively.

  Then, the band-by-band suppression unit 102 performs noise component suppression processing for each band on the input spectra IS1 to ISM converted into the frequency domain, and the post-suppression spectrum OS1 to OS1 as signals after noise component suppression (frequency domain signals). Create an OSM.

  In the band-by-band suppression unit 102, the channel selection unit 201 selects channel numbers to be cyclically applied to the selected input spectrum in the order of channel numbers. The suppression gain calculation unit 202 generates a suppression gain based on the selected input spectrum selected by the channel selection unit 201. In the band-by-band suppression unit 102, the multiplication units 203-1 to 203-M perform a process of multiplying the input spectra IS1 to ISM by a common suppression gain (a suppression gain supplied from the suppression gain calculation unit 202). Is called. Thereby, the multiplication means 203-1 to 203-M suppress the noise components included in the input spectra IS1 to ISM, respectively, and output the suppressed spectra OS1 to OSM.

  Then, the waveform restoration units 103-1 to 103 -M convert the suppressed spectra OS <b> 1 to OSM supplied from the band-based suppression units 102 into time-domain signals, and output them as post-suppression signals os <b> 1 to osM.

(A-3) Effects of First Embodiment According to the first embodiment, the following effects can be achieved.

  In the noise suppression apparatus 100 of the first embodiment, the amount of channel number selection processing applied to the selected input spectrum by the channel selection unit 201 is constant even when the number of channels of the input signal increases. Therefore, in the noise suppression apparatus 100 according to the first embodiment, even when the number of channels of the input signal is increased, the order of O (n) is not increased according to the number of channels, and noise suppression is performed with a smaller processing amount than in the past. It can be performed.

  Further, in the noise suppression device 100 of the first embodiment, a plurality of suppression gains are generated by switching the selection input spectrum by an arbitrary selection method (for example, a method of cyclically selecting in the order of channel numbers). Noise suppression can be performed by generating a suppression gain that reflects the average input power spectrum of the channel.

  Hereinafter, in the noise suppression apparatus 100, it is possible to maintain the noise suppression performance even when generating the suppression gain while switching the selected input spectrum by an arbitrary selection method (for example, a method of cyclically selecting in order of channel numbers). Explain the mathematical support of this.

  As the selection method for selecting a channel number to be applied to the selected input spectrum in the channel selection means 201, for example, the following three selection methods can be mentioned.

  The first selection method is a method in which the channel selection unit 201 cyclically selects channel numbers in order of channel number for each unit time for which the suppression gain calculation unit 202 executes suppression gain calculation (the first embodiment described above). Is the same selection method). In the first selection method, the channel selection unit 201 selects, for example, channel 1, channel 2,..., And the next channel M is selected, as in the first embodiment. After the unit time, channel 1 is selected again.

  In the second selection method, the channel selection unit 201 uses a predetermined order (for example, an order based on a preset list) for each unit time for which the suppression gain calculation unit 202 executes the suppression gain calculation. This is a method of selecting a number (the same selection method as in the second embodiment described later).

  The third selection method is a method in which the channel selection unit 201 selects channel numbers in an order based on pseudo-random numbers (random order) (selection method similar to the third embodiment described later). Specifically, the channel selection unit 201 generates a pseudo-random number, compares the pseudo-random number with a threshold value that is one less than the predetermined number of channels, and sets a threshold value lower than the pseudo-random number. A channel number corresponding to the number may be selected.

When the channel selection unit 201 selects channel numbers 1 to M (M = 4) by the above-described first to third selection methods, the processing amount order is the first and second selection methods. Since O (1) and O (log ₂ n) in the third selection method, the processing order for calculating the suppression gain in the noise suppression apparatus 100 is also O (1) or O (log ₂ n). . That is, in this case, the noise suppression apparatus 100 can realize suppression gain calculation with a processing amount that does not depend on the number of channels of the input signal.

  Since the first selection method and the second selection method described above are the same as selecting a channel stochastically in the long term (for example, several seconds or more), the third selection method described above It can be regarded as a special case. Therefore, in the following, the reason why the suppression performance can be maintained in the noise suppression device 100 will be described using the example of the third selection method described above.

  In order for the suppression gain calculation means 202 to calculate the suppression gain, the noise power spectrum must first be estimated. When the input signal is a single channel, generally, the k-th element (synonymous with “k-th band”) N (k, t) of the noise power spectrum at time t is the element X (k, t) of the input power spectrum. ) And the noise weighting factor W (k, t) as an expected value is estimated by the following equation (4). In the following expression (4), operators E and {} mean expected values for time t.

In the following equation (4), the noise weighting coefficient W (k, t) is 0 if the input power spectrum element is completely noisy, and is 1 if it contains a lot of speech. In the following equation (4), the speech interval may be detected and the noise weighting coefficient may be a discrete value of 0 or 1. If the speech interval or the noise interval is warm, the noise weighting coefficient is 0 or more and 1 The following real values may be used.
N (k, t) = E _t {W (k, t) · X (k, t)} (4)

Here, a method for generating a representative signal from a plurality of channels and estimating a noise power spectrum using the representative signal is formulated. Since generation of a representative signal by averaging is a special case of weighted addition, a method of generating a representative signal by weighted addition is applied here. Then, assuming that the channel weight coefficient for the input power spectrum X _m (k, t) of the channel _m is C _m , the noise power spectrum estimation formula is the following formula (5). However, _{C m} shall meet the following (6).

If the above equation (6) is satisfied, the channel weighting coefficient C _m is regarded as a probability distribution. Therefore, here, a channel selection factor B _m (t) that satisfies the following expression (7) is introduced. The channel selection factor B _m (t) is such that only one channel number is 1 at a certain time t and all other channel numbers are 0, but the channel m that is 1 is randomly selected according to the probability distribution C _m. It is assumed that these are factors. In order to introduce such a channel selection factor B _m (t), first, the expected value of the following equation (5) and the calculation order of the sum related to the channel are changed and transformed as the following equation (8). . Then, when the probability distribution C _m in the following equation (8) is rewritten with the expected value of the channel selection factor B _m (t), the following equation (9) is obtained. Since the channel selection factor B _m (t) is independent of the input power spectrum X _m (k, t) and the noise weighting factor W (k, t), the channel selection factor B _m (t) can be transformed as the following equation (10), and finally again The following equation (11) can be obtained by changing the calculation order of the sum of the expected value and the channel.

Then, since the left side of the above equation (5) is equal to the following equation (11), the channel selection factor B _m (t) is multiplied by the input power spectrum X _m (k, t) to obtain the sum relating to the channel. It can be seen that this means that one channel is selected at random according to the probability distribution C _m from the input power spectrum X _m (k, t).

Therefore, estimating a noise power spectrum by randomly selecting one channel from the input power spectrum X _m (k, t) according to the probability distribution C _m is represented by weighted addition based on the channel weight coefficient C _m. It can be seen that this is the same as generating a signal and estimating the noise power spectrum.

  As described above, in the noise suppression apparatus 100, it is possible to maintain the noise suppression performance even when the suppression gain is generated while switching the selected input spectrum by any of the first to third selection methods described above. Mathematically supported. That is, in the noise suppression apparatus 100, it can be said that the processing amount order for calculating the suppression gain can be reduced to less than O (n) and that the suppression performance can be maintained even if the processing amount is reduced.

(B) Second Embodiment Hereinafter, a second embodiment of the noise suppression device, program and method according to the present invention will be described in detail with reference to the drawings.

(B-1) Configuration and Operation of Second Embodiment The overall configuration of the noise suppression device 100A of the second embodiment can also be described using FIG. Hereinafter, differences of the second embodiment from the first embodiment will be described.

  In the noise suppression apparatus 100 according to the first embodiment, a suppression gain is generated by selecting a channel number to be applied to a selected input spectrum by the above-described first selection method (a method of cyclically selecting channel numbers in order of channel numbers). Was. On the other hand, in the noise suppression apparatus 100A of the second embodiment, the channel number applied to the selected input spectrum ISS is selected by the above-described second selection method (method of selecting channel numbers in a predetermined order). Select to generate suppression gain.

  As shown in FIGS. 1 and 2, the noise suppression apparatus 100A is the first implementation in that the band-by-band suppression unit 102 (channel selection unit 201) is replaced with the band-by-band suppression unit 102A (channel selection unit 201A). It is different from the form.

  FIG. 6 is a block diagram showing an internal configuration of the channel selection unit 201A according to the second embodiment.

  As shown in FIG. 6, the channel selection unit 201A is different from the first embodiment in that the channel determination unit 302 is replaced with the channel determination unit 302A, and the channel storage unit 303 is excluded.

  The channel determining unit 302A determines a channel number based on the frame number given from the frame counting unit 301, and gives the obtained channel number to the input spectrum selecting unit 304 as a selected channel number.

  FIG. 7 is a block diagram showing the internal configuration of the channel determination unit 302A of the second embodiment.

  As shown in FIG. 7, the channel determination unit 302A includes a channel order supply unit 601 and a channel order selection unit 602.

  The frame number given to the channel determining means 302A is given to the channel order selecting means 602.

  The channel order supply unit 601 gives a list (hereinafter referred to as “channel order information”) describing a preset order of channel numbers to the channel order selection unit 602. The channel order information can be in an arbitrary format such as an array, a vector, or a list, for example, but is preferably defined by a one-way circulation list. In the following, it is assumed that the channel order information is defined by a one-way circulation list. The elements of the channel order information are limited to any one of channel numbers 1 to M, but the number of elements of the channel order is arbitrary. However, when the number of elements described in the channel order information is less than M, a channel that is never referenced is generated. In addition, when the number of elements of the channel order information is exactly M, the appearance probability of each channel number is statistically the same as in the first embodiment. Therefore, it is preferable to set more than M elements to be set in the channel order information. For example, when M = 4, if channel order information is defined as {2, 2, 2, 3, 3, 1, 4}, channel number 2 is most important, then channel number 3 is most important, and channel number 1 and 4 are weighted so that they are referenced but not regarded as important. Also, for example, as the channel order information, the order of {2, 2, 2, 3, 3, 1, 4} is changed and set to {1, 2, 3, 4, 2, 3, 2} Also good.

  The channel order selection means 602 selects one channel from the channel order based on the frame number and sets it as the selected channel number. The channel order selection means 602 determines in advance the channel order to be selected first as the initial channel order. If the frame number is the initial frame number, the initial channel order is set as the selected channel number. The channel order is the selected channel number.

  The initial frame number supplied by the frame counting means 301 varies depending on the implementation, but may be set to 0 or 1, for example.

(B-2) Effects of Second Embodiment According to the second embodiment, the following effects can be achieved.

  In the noise suppression device 100A of the second embodiment, as in the first embodiment, even when the number of channels of the input signal increases, the processing amount order of O (n) does not become an order according to the number of channels. Noise suppression can be performed with a smaller amount of processing than before.

  Further, in the noise suppression device 100A of the second embodiment, the appearance probability can be weighted for each channel number with the description content of the list of channel order information. For example, in the noise suppression device 100A of the second embodiment, it is possible to suppress noise while evaluating the input power spectra of a plurality of channels according to a predetermined importance degree. For example, a more accurate suppression gain can be obtained by describing a list of channel order information so that the appearance frequency of the channel number related to the input signal captured by the microphone closest to the position of the noise source is increased. Can do.

(C) Third Embodiment Hereinafter, a third embodiment of the noise suppression device, program and method according to the present invention will be described in detail with reference to the drawings.

(C-1) Configuration and Operation of Third Embodiment The overall configuration of the noise suppression device 100B of the third embodiment can also be described using FIG. 1 described above. Hereinafter, differences of the third embodiment from the third embodiment will be described.

  In the noise suppression apparatus 100 according to the first embodiment, a suppression gain is generated by selecting a channel number to be applied to a selected input spectrum by the above-described first selection method (a method of cyclically selecting channel numbers in order of channel numbers). Was. On the other hand, in the noise suppression apparatus 100 of the third embodiment, the channel number to be applied to the selected input spectrum is selected by the above-described third selection method (a method of selecting channel numbers in the order based on pseudo-random numbers). To generate a suppression gain.

  As shown in FIGS. 1 and 2, the noise suppression apparatus 100B is the first implementation in that the band-by-band suppression unit 102 (channel selection unit 201) is replaced with the band-by-band suppression unit 102B (channel selection unit 201B). It is different from the form.

  FIG. 8 is a block diagram showing the internal configuration of the channel selection unit 201B according to the third embodiment.

  As shown in FIG. 8, the channel selection unit 201B is different from the first embodiment in that the channel determination unit 302 is replaced with the channel determination unit 302B, and the frame counting unit 301 and the channel storage unit 303 are omitted. ing.

  The channel determination unit 302B determines a channel number based on the pseudo random number, and gives the obtained channel number to the input spectrum selection unit 304 as a selected channel number.

  FIG. 9 is a block diagram showing the internal configuration of the channel determination unit 302B of the third embodiment.

  As illustrated in FIG. 9, the channel determination unit 302B includes a random number generation unit 701, a threshold set supply unit 702, and a threshold determination unit 703.

  The random number generation unit 701 generates a pseudo random number r and gives it to the threshold value determination unit 703. The method used by the random number generation unit 701 for generating pseudo-random numbers is not limited. For example, it is desirable to apply a linear congruential method with a small processing amount. Further, in the random number generation unit 701, for example, if the processing amount and the resource are not problematic, it is desirable to apply, for example, a Mersenne twister having a longer period and less regularity.

  Since only an integer value is used for the calculation of the linear congruential method, the prime pseudorandom number obtained is an integer value (referred to as an integer pseudorandom number). A pseudo-random number r obtained by dividing the random number by the maximum value of the integer pseudo-random number and converting it to a real value of 0.0 to 1.0 is output.

The threshold set supply unit 702 predefines a set of real values greater than 0.0 and less than 1.0 (M−1), which is 1 less than the number of channels M, and sets the defined set as the threshold set R. _i (i = 1,..., M−1) is given to the threshold determination means 703. The configuration of the threshold set is not limited, but is preferably a balanced binary search tree using an array (so that the elements of the array satisfy R _2i <R _i <... <R _{2i + 1} ). (FIG. 10). “Array” in FIG. 10 indicates that threshold values are stored in the array in the order of the subscript i. Similarly, “data structure” indicates that the threshold set has a structure of an equilibrium binary search tree. Similarly, “value” indicates an image of where (M−1) threshold values are located on a number line in the range of 0.0 to 1.0. In FIG. 10, the threshold values are located at equal intervals, but the threshold value is not limited to this. Similarly, “selected channel” indicates the relationship between the position of the random number on the number line and the selected channel number selected by the threshold determination means 703 described later.

The threshold determination unit 703 sets a channel number determined based on the given pseudorandom number r and the given threshold set R _i as the selected channel number. The selected channel number m is selected according to the flow shown in FIG. The operation of the threshold determination unit 703 according to FIG. 11 will be described. First, initialization is performed by substituting the channel number M for the selected channel number m (S1), and a value (0.0 to 1....) Generated by the random number generation means 701 (corresponding to the function rand () in FIG. 11) for the pseudorandom number r. (Real value of 0) is substituted (S2), and 1 is substituted for the subscript i of the threshold set for initialization (S3). Compares the pseudo-random number r and the threshold value _{R i} (S4), by substituting r _{<R i} is long if the selected channel number m subscript i (S5), and update doubled i (S6), If r ≧ R _i , i is doubled and 1 is added to update (S7). Then, the subscript i is compared with the number of channels M (S8). If i <M, S4 to S7 are updated, and if i ≧ M, the process is terminated and the selected channel number m is determined. According to the above method, the number of comparisons between the pseudo random number and the threshold value is an integer number obtained by rounding up log ₂ M at most, that is, the processing order O (log ₂ n), which is smaller than O (n) in Patent Document 1. Can be realized.

Therefore, in the channel determination unit 302B of the third embodiment, which channel is to be regarded as important (appearance rate of each channel number) depending on how the threshold value array R _i is determined (range applied to each channel number). Can be adjusted. For example, if R _i = i / M, the appearance rates of the respective channel numbers are equal, and the appearance rates of the respective channel numbers are statistically the same as those in the first embodiment. Further, by adjusting the range corresponding to the channel number to be regarded as important, it is possible to make adjustments such as increasing the appearance rate of the channel numbers regarded as important statistically, as in the second embodiment.

(C-2) Effects of Third Embodiment According to the third embodiment, the following effects can be achieved.

In the noise suppression device 100B of the third embodiment, as in the first and second embodiments, even when the number of channels of the input signal increases, the processing amount order of O (n) according to the number of channels In addition, noise suppression can be performed with a smaller processing amount than in the past. In the noise suppression device 100B according to the third embodiment, random number generation is executed once per unit time, and threshold determination is executed only by an integer number obtained by rounding up log ₂ M at most, so that the processing amount order is a balanced binary search tree. The same O (log ₂ n).

Further, in the third embodiment, by adjusting the contents of the threshold array R _i (i = 1,..., M−1) set in the channel determining unit 302B (adjusting the range corresponding to each channel number), Noise can be suppressed while evaluating the input power spectrum of a plurality of channels according to a predetermined degree of importance with a high degree of freedom.

  In the first embodiment and the second embodiment, the order of selecting channels is uniquely determined. In the first embodiment, only a plurality of channels can be referred to on average. In the second embodiment, it is not easy to freely set the degree of importance of each channel. For example, in the second embodiment, if the number of channels is M = 2, and channel 1 is 30% and channel 2 is 70% important, {1, 1, 1, 2, 2 for only 2 channels , 2, 2, 2, 2, 2} and the like, a channel order having 10 or more elements must be prepared. However, the noise suppression apparatus 100B of the third embodiment generates a pseudo random number, and determines a channel based on the generated pseudo random number and a predetermined threshold value. This makes it possible to easily realize settings that are difficult in the second embodiment, such as 13% and 87%, for example, when the number of channels M = 2. That is, in the third embodiment, the appearance rate of each channel number can be adjusted more easily than in the second embodiment.

(D) Other Embodiments The present invention is not limited to the above-described embodiments, and may include modified embodiments as exemplified below.

In (D-1) second and third embodiments, a user may be capable of changing the channel order information and threshold array R _i (Edit).

  In the second embodiment, when the user can change the degree of importance on the way, channel order automatic generation means for automatically generating the channel order according to the degree of importance may be provided. It is desirable that this automatic channel order generation means, for example, determines the maximum number of elements of the channel order in advance and configures the channel order so that the degree of importance designated by the user is reflected as accurately as possible.

  Furthermore, in the third embodiment, when the user can change the degree of importance on the way, a threshold value array generating unit configured to automatically adjust the threshold value array in response to a user operation instead of the threshold value array supplying unit is configured. It is desirable to include.

(D-2) In the second and third embodiments, channel importance adapting means for automatically changing the degree of importance according to the input power spectrum of a plurality of channels may be included in the configuration. For example, the channel importance level adaptation unit may be configured to change the degree of importance so that the channel having the maximum audio power index value of each channel is more important. As the audio power index value, for example, it is desirable to apply the maximum value of audio power for a certain period. The audio power index value may be updated only on the channel of the selected channel number at a certain time. For example, the channel importance adapting means includes a value obtained by multiplying a voice power index value one unit time before by a forgetting factor β (for example, β = 0.99) and a maximum element of the input power spectrum of the selected channel number. The larger one may be updated as a new voice power index value. The channel importance adapting means stores, for example, the maximum audio power index value among the audio power index values of a plurality of channels one unit time before, and the audio power index value of the selected channel number is the maximum The degree of importance may be changed so that the selected channel number is more important if the value is larger than the voice power index value. In this way, it is possible to change the degree of emphasis while maintaining the processing amount order at O (log ₂ n).

  101, 101-1 to 101-M ... frequency analysis means, 102 ... band-based suppression means, 103-1 to 103-M ... waveform restoration means, 201 ... channel selection means, 202 ... suppression gain calculation means, 203-1 203-M... Multiplying means 301... Frame counting means 302. Channel determining means 303... Channel storage means 304... Input spectrum selecting means 401. Switch means 501... Power calculation means 502... Noise smoothing means 503. Noise storage means 504 .. Post-SNR calculation means 505 .. Post-SNR storage means 506... Threshold determination means 507. Prior SNR estimation means, 509 ... suppression gain determination means, 510 ... suppression gain storage means.

Claims (7)

In a noise suppression device that suppresses noise from an input spectrum corresponding to each of multiple channel input signals,
Channel selection means for selecting one channel from the plurality of channels by a predetermined method every predetermined unit time;
Suppression gain calculating means for calculating a suppression gain for suppressing a noise component included in the input spectrum using the input spectrum of the selected channel;
A noise suppression device comprising: noise suppression means for suppressing a noise component for each of the input spectra of the plurality of channels using the suppression gain. Channel numbers are assigned to the channels constituting the plurality of channels,
The noise suppression apparatus according to claim 1, wherein the channel selection means performs channel selection cyclically in the order of channel numbers from the plurality of channels.   2. The channel selection unit according to claim 1, wherein the channel selection unit selects one of the plurality of channels in an order according to channel order information in which an order of selecting a channel from the plurality of channels is described. Noise suppression device. A pseudo random number generating means for generating a pseudo random number;
The noise suppression apparatus according to claim 1, wherein the channel selection unit selects one of the plurality of channels based on the pseudorandom number generated by the pseudorandom number generation unit. Further comprising threshold value array means for holding a threshold value array composed of threshold values that are one less than the number of channels of the plurality of channels;
5. The noise suppression according to claim 4, wherein the channel selection unit selects a channel based on a comparison result between the pseudo-random number generated by the pseudo-random number generation unit and the threshold value constituting the threshold value array. apparatus. A computer mounted in a noise suppression device that suppresses noise from the input spectrum corresponding to each of the input signals of multiple channels,
Channel selection means for selecting one channel from the plurality of channels by a predetermined method every predetermined unit time;
Suppression gain calculating means for calculating a suppression gain for suppressing a noise component included in the input spectrum using the input spectrum of the selected channel;
A noise suppression program that functions as a noise suppression unit that suppresses a noise component for each of the input spectra of the plurality of channels using the suppression gain. In a noise suppression method for suppressing noise from an input spectrum corresponding to each of input signals of a plurality of channels,
A channel selection means, a suppression gain calculation means, and a noise suppression means,
The channel selection means selects one channel from the plurality of channels by a predetermined method every predetermined unit time,
The suppression gain calculation means calculates a suppression gain for suppressing a noise component included in the input spectrum using the input spectrum of the selected channel,
The noise suppression method, wherein the noise suppression means suppresses a noise component for each of the input spectra of the plurality of channels using the suppression gain.

Images