JP2018142826A - Non-target sound suppression device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To achieve the satisfactory sound quality of target sound, and to control a suppression coefficient or subtraction coefficient with a low processing load when reducing or subtracting non-target sound from input signals.SOLUTION: A non-target sound suppression device according to the present invention, comprises: a front suppression signal generation unit which generates front suppression signals with a blind spot in front thereof based on differences among a plurality of frequency domain input signals resulting from conversion of input signals from a time domain to a frequency domain; a coherence calculating unit which calculates a coherence based on signals obtained from the plurality of input signals; a feature quantity calculating unit which calculates a feature quantity indicative of the relationship between the front suppression signal and the coherence; and a non-target sound suppression processing unit which uses the feature quantity indicative of the relationship between the front suppression signal and the coherence to set a coefficient involved in suppression of non-target sound included in the input signals, and obtains post-suppression processing signals produced by using the coefficient to suppress the non-target sound included in the input signals.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a non-target sound suppressing apparatus, method, and program, and can be applied to, for example, a communication device or communication software that uses voice, such as a telephone or a video conference system, or an acoustic signal process used in a preprocessing of voice recognition It is.

  In recent years, devices equipped with various voice processing functions such as a voice call function and a voice recognition function have become widespread, such as smartphones and car navigation systems. However, with the widespread use of these devices, voice processing functions are being used in harsher noise environments than before, such as in crowded streets and running cars. For this reason, there is an increasing demand for signal processing technology that can maintain call sound quality and speech recognition performance even in a noisy environment.

  Noise that hinders the performance of the voice processing function includes, for example, background noises such as crowds in the streets and driving noise of automobiles, and disturbing sounds (for example, disturbing voices such as the speech of people other than the user of the voice processing function) Can be broadly divided. Various effective suppression methods have been proposed on the assumption that the background noise has a steady frequency characteristic and power (see Patent Documents 1 to 3 and Non-Patent Document 1).

Japanese translation of PCT publication 2010-532879 JP 2014-106337 A JP 2014-164191 A

Published by Kazuyuki Hiraoka and Gen Hori, “Probability Statistics for Programming,” Ohmsha, October 23, 2009

  However, as described above, the background noise is not steady due to the rapid expansion of the usage environment of the audio signal processing function. Therefore, there is a need for a background noise suppression method that can quickly follow fluctuations in the characteristics of background noise.However, when background noise is suppressed in a signal section in which an interfering sound exists, the signal component of the target sound is also lost and the sound quality is improved. Degradation may occur.

  Patent Document 3 discloses a technique for suppressing interfering sounds coming from the surroundings by subtracting a signal (referred to as a front suppression signal) in which a component coming from the front is suppressed from the input signal. When subtracting, the intensity of subtraction is often controlled by multiplying the front suppression signal by the subtraction coefficient.If the subtraction coefficient is too large, the suppression performance is excessive and distortion of the target sound increases. It has a great influence on sound quality such as insufficient suppression performance. However, it is difficult to determine the presence of the interference sound superimposed on the target sound, and it is difficult to set the subtraction coefficient to an appropriate value.

  Therefore, in view of the above problems, when suppressing or subtracting a non-target sound from an input signal, the non-target sound suppressing apparatus can improve the sound quality of the target sound, suppress the processing load, and control the suppression coefficient or the subtraction coefficient. There is a need for methods and programs.

  In order to solve such a problem, the non-target sound suppression apparatus according to the first aspect of the present invention is (1) a plurality of input signals obtained by converting each input signal from each of a plurality of microphones from a time domain to a frequency domain. A front suppression signal generation unit that generates a frontal suppression signal having a blind spot in front based on a difference in frequency domain input signals; and (2) a coherence calculation unit that calculates coherence based on signals obtained from a plurality of input signals; (3) a feature value calculation unit for calculating a feature value indicating the relationship between the front suppression signal and coherence; and (4) a feature value indicating the relationship between the front suppression signal and coherence. A non-target sound suppression processing unit that sets a coefficient related to suppression of the non-target sound and obtains a post-suppression signal that suppresses the non-target sound included in the input signal using the coefficient. .

  The non-target sound suppression method according to the second aspect of the present invention includes: (1) a plurality of frequencies obtained by the front suppression signal generation unit converting each input signal from each of the plurality of microphones from the time domain to the frequency domain; Based on the difference between the region input signals, a front suppression signal having a blind spot in front is generated. (2) The coherence calculation unit calculates coherence based on signals obtained from a plurality of input signals. The calculation unit calculates a feature amount indicating the relationship between the front suppression signal and the coherence, and (4) the non-target sound suppression processing unit uses the feature amount indicating the relationship between the front suppression signal and the coherence. A coefficient relating to suppression of the non-target sound included in the signal is set, and a signal after suppression processing in which the non-target sound included in the input signal is suppressed is obtained using the coefficient.

  A non-target sound suppression program according to a third aspect of the present invention is a computer program comprising: (1) a plurality of frequency domain input signals obtained by converting each input signal from each of a plurality of microphones from a time domain to a frequency domain; A front suppression signal generation unit that generates a frontal suppression signal having a blind spot in front based on the difference; (2) a coherence calculation unit that calculates coherence based on signals obtained from a plurality of input signals; (4) a feature amount calculation unit that calculates a feature amount indicating a relationship between the suppression signal and coherence; and (4) a feature amount indicating a relationship between the front suppression signal and the coherence. A coefficient related to suppression is set, and the coefficient is used as a non-target sound suppression processing unit that obtains a signal after suppression processing in which the non-target sound included in the input signal is suppressed using the coefficient.

  According to the present invention, when suppressing or subtracting a non-target sound from an input signal, it is possible to control the suppression coefficient or the subtraction coefficient with good sound quality of the target sound and low processing load.

It is a block diagram which shows the whole structure of the non-target sound suppression apparatus which concerns on 1st Embodiment. It is explanatory drawing explaining the example of arrangement | positioning of the microphone which concerns on embodiment. It is a figure which shows the characteristic of the directional signal applied with the acoustic signal processing apparatus which concerns on embodiment. It is a block diagram which shows the structure of the WF part which concerns on 1st Embodiment. It is a flowchart which shows the process in the time constant control part of the WF part which concerns on 1st Embodiment. It is a block diagram which shows the whole structure of the non-target sound suppression apparatus which concerns on 2nd Embodiment. It is a block diagram which shows the structure of the frequency subtraction process part which concerns on 2nd Embodiment. It is a flowchart which shows the process in the time constant control part 23 of the frequency subtraction process part which concerns on 2nd Embodiment.

(A) First Embodiment Hereinafter, a first embodiment of a non-target sound suppressing apparatus, method and program according to the present invention will be described in detail with reference to the drawings.

  In the first embodiment, by using the present invention, a background noise suppression apparatus and method (non-target sound suppression apparatus) that quickly follows fluctuations in characteristics of non-steady background noise due to rapid expansion of the usage environment of the audio signal processing function. And method).

  Here, when the background noise suppression function is used in an environment where interference sound is generated in the surroundings, the coefficient adaptation operation may be erroneously performed in a signal section where the interference sound exists. At this time, the characteristic of human voice as an interfering sound is also reflected in the background noise suppression coefficient (hereinafter referred to as “suppression coefficient”), and therefore when the suppression process is performed using the coefficient, the signal of the target sound The component may be lost, and the sound quality may be deteriorated.

  Therefore, in the first embodiment, in order to prevent the above phenomenon, the background noise fluctuation is continuously monitored while suppressing the influence of the target sound and the interference sound, and the background noise suppression coefficient adaptive operation is performed based on the result. A non-target sound suppressing apparatus and method capable of controlling the sound is realized.

(A-1) Configuration of the First Embodiment FIG. 1 is a block diagram showing the overall configuration of the non-target sound suppression apparatus 1 according to the first embodiment.

  As shown in FIG. 1, the non-target sound suppression apparatus 1 acquires input signals s1 (n) and s2 (n) from a plurality of microphones m_1 and m_2 (two cases are shown in FIG. 1). . Note that n is an index indicating the input order of samples and is expressed by a positive integer. In the following, it is assumed that the smaller n is the older input sample, and the larger n is the new input sample.

  The non-target sound suppression apparatus 1 sets a parameter (variable) that suppresses background noise by following changes in characteristics of the background noise based on each input signal acquired from the microphones m_1 and m_2, and suppresses the background noise. The post-suppression signal is supplied to the subsequent audio processing device 2.

  The audio processing device 2 performs predetermined audio processing using the post-suppression signal from the non-target sound suppressing device 1. The processing contents in the voice processing device 2 are not particularly limited, and those that perform various processing can be applied. For example, voice communication processing or voice recognition processing such as a telephone terminal or a video conference system may be performed. Good. It should be noted that the non-target sound suppressing device 1 and the sound processing device 2 may be configured so as to be able to exchange signals, and may be connected to a circuit wiring, for example, via a wired line or a wireless line. It may be one that can send and receive signals by network communication.

  FIG. 2 is an explanatory diagram for explaining an arrangement example of the microphones m_1 and m_2.

  As shown in FIG. 2, the microphones m_1 and m_2 are arranged so that the plane including the two microphones m_1 and m_2 is perpendicular to the direction in which the target sound arrives (the direction of the target sound source). And In the following, as shown in FIG. 2, the arrival direction of the target sound is referred to as the front direction or the front direction when viewed from the position between the two microphones m_1 and m_2. In the following, as shown in FIG. 2, when referring to the right direction, the left direction, and the backward direction, each direction when viewing the arrival direction of the target sound from the position between the two microphones m_1 and m_2 is shown. It will be explained as a thing. In this embodiment, it is assumed that the target sound comes from the front direction of the microphones m_1 and m_2, and the non-target sound including the interference sound comes from the left-right direction (lateral direction).

  As shown in FIG. 1, the non-target sound suppression apparatus 1 includes an FFT unit 11, a front suppression signal generation unit 12, a coherence calculation unit 13, a correlation and modGI calculation unit 14, a WF (Wiener filter) unit 15, and an IFFT unit 16. Have.

  The non-target sound suppression device 1 may be realized by installing a program (for example, a non-target sound suppression program) in a computer having a processor, a memory, and the like. In this case, the non-target sound suppression device 1 is functional. Can be shown using FIG. Note that part or all of the non-target sound suppressing device 1 may be realized by hardware.

  The FFT unit 11 receives the input signals s1 and s2 from the microphones m_1 and m_2 via an AD converter (not shown), and performs fast Fourier transform (or discrete Fourier transform) on the input signals s1 and s2. . As a result, the input signals s1 and s2 are expressed in the frequency domain.

  Note that, in performing the fast Fourier transform, the FFT unit 11 includes an analysis Fourier FRAME1 (K) and a predetermined N (N is an arbitrary integer) samples from the input signals s1 (n) and s2 (n). Assume that FRAME2 (K) is configured. An example of configuring FRAME1 from the input signal s1 is shown in the following equation (1).

  In the equation (1), K is an index indicating the order of frames and is expressed by a positive integer. In the following, it is assumed that the smaller the K value, the older the analysis frame, and the larger the K value, the newer the analysis frame. In the following description, it is assumed that the index representing the latest analysis frame to be analyzed is K unless otherwise specified.

  The FFT unit 11 performs a fast Fourier transform process for each analysis frame, thereby performing a frequency domain signal X1 (f, K) obtained by performing a Fourier transform on the analysis frame FRAME1 (K) configured from the input signal s1, and an input A frequency domain signal X2 (f, X) obtained by performing Fourier transform on the analysis frame FRAME2 (K) configured from the signal s2 is supplied to the front suppression signal generation unit 12 and the coherence calculation unit 13.

  Here, f is an index representing a frequency. Further, the frequency domain signal X1 (f, K) is not a single value but is composed of m (m is an arbitrary integer) spectral components of a plurality of frequencies f1 to fm as shown in the equation (2). Suppose it is a thing.

  In the above equation (2), X1 (f, K) is a complex number and consists of a real part and an imaginary part. The same applies to X2 (f, K) and the front suppression signal N (f, K) described in the front suppression signal generation unit 12 described later.

  The front suppression signal generation unit 12 performs a process of suppressing the signal component in the front direction for each frequency with respect to the signal supplied from the FFT unit 11. In other words, the front suppression signal generation unit 12 functions as a directivity filter that suppresses a component in the front direction.

  For example, as shown in FIG. 3, the front suppression signal generation unit 12 uses an 8-shaped bi-directional filter having a blind spot in the front direction to generate a front direction component from the signal supplied from the FFT unit 11. A directional filter that suppresses the noise is formed.

Specifically, the front suppression signal generation unit 12 performs a calculation such as the following equation (3) based on the signals X1 (f, K) and X2 (f, K) supplied from the FFT unit 11. Thus, the front suppression signal N (f, K) for each frequency is generated. The calculation of the following equation (3) corresponds to a process of forming an 8-shaped bi-directional filter having a blind spot in the front direction as shown in FIG.
N (f, K) = X1 (f, K) -X2 (f, K) (3)

  As described above, the front suppression signal generation unit 12 acquires each frequency component of frequencies f1 to fm (power for one frame in each frequency band).

  Further, the front suppression signal generator 12 calculates an average front suppression signal AVE_N (K) by averaging the front suppression signals N (f, K) over all frequencies f1 to fm according to the equation (4). To do.

  The coherence calculation unit 13 forms a highly directional signal in a specific direction included in the frequency domain signals X1 (f, K) and X2 (f, K) from the FFT unit 11 to calculate coherence COH (K). .

  Here, the calculation processing of coherence COH (K) in the coherence calculation unit 13 will be described.

  The coherence calculator 13 processes the signal B1 (f, K) obtained by processing the frequency domain signals X1 (f, K) and X2 (f, K) with a filter having strong directivity in the first direction (for example, the left direction). The coherence calculation unit 13 forms the signal B2 (f) processed from the frequency domain signals X1 (f, K) and X2 (f, K) with a filter having strong directivity in the second direction (for example, the right direction). , K). An existing method can be applied to the formation method of the signals B1 (f) and B2 (f) having high directivity in a specific direction. Here, the following equation (5) is applied to the first direction. An example in which the signal B1 having high directivity is formed and the signal B2 having high directivity in the second direction is formed by applying the following equation (6) will be described.

  In the above formulas (5) and (6), S is the sampling frequency, N is the FFT analysis frame length, τ is the difference in arrival time of sound waves between the microphone m_1 and the microphone m_2, i is the imaginary unit, and f is the frequency. .

Next, the coherence calculation unit 13 performs the operations shown in the following expressions (7) and (8) on the signals B1 (f) and B2 (f) obtained as described above. Obtain coherence COH (K). Here, B2 (f, K) ^* in the equation (7) is a conjugate complex number of B2 (f, K).

  coef (f, K) is a coherence in a component of an index K having an arbitrary index K (an arbitrary frequency f (any one of frequencies f1 to fm) constituting the analysis frames FRAME1 (K) and FRAME2 (K)). .

  When obtaining coef (f, K), if the directionality of the signal B1 (f) is different from that of the signal B (f), the signals B1 (f) and B2 ( The directivity direction according to f) may be any direction other than the front direction. Further, the method for calculating coef (f, K) is not limited to the above calculation method.

  The correlation and mod GI calculation unit 14 obtains a front suppression signal N (f, N) (average front suppression signal AVE_N (K)) and coherence COH (K) having directivity other than the front, and an average front suppression signal. A correlation coefficient cor (K), which is a feature amount indicating the relationship between AVE_N (K) and coherence COH (K), is calculated.

  Further, the correlation and mod GI calculation unit 14 uses the correlation coefficient cor (K) to represent a feature amount (cor_modGI (K)) that represents the intensity of positive and negative fluctuations in the amplitude slope of the correlation coefficient cor (K). And the characteristic amount (cor_modGI (k)) is output to the WF unit 15.

  First, a description will be given of the principle in which the correlation and ModGI calculation unit 14 detects a signal section in which an interfering sound exists based on the correlation coefficient cor (K) between the average front suppression signal AVE_N (K) and the coherence COH (K). To do.

  Here, there is a sound source that emits a target sound in the front direction of the microphone m_1 and the microphone m_2, and the disturbing sound is generated from a direction other than the front direction (for example, the lateral direction of the microphone m_1 and the microphone m_2 (that is, the left direction and the right direction). Shall arrive.

  For example, when “no disturbing voice is present” and “the target sound is present”, the front suppression signal N (f, K) has a signal value proportional to the magnitude of the target sound component. However, as shown in FIG. 2, since the gain in the front direction is smaller than the gain in the horizontal direction, the gain is smaller than that in the case where an interfering sound is present.

  The coherence COH (K) is a feature quantity that has a deep relationship with the arrival direction of the input signal, and is rephrased as a correlation between two signal components. This is because the equation (6) is an equation for calculating the correlation for a certain frequency component, and the equation (7) is an equation for calculating the average of the correlation values of all frequency components, so that the coherence COH ( When K) is small, it can be said that the correlation between the two signal components is small, and conversely, when the coherence COH (K) is large, it can be said that the correlation between the two signal components is large. The input signal when the coherence COH (K) is small can be said to be a signal arriving from a direction other than the front direction because the arrival direction is greatly biased to either the right direction or the left direction. On the other hand, the input signal when the coherence COH (K) is large can be said to be a signal arriving from the front direction with little deviation in the arrival direction.

  Then, if “no interference sound exists” and “the target sound exists”, the coherence COH (K) becomes a large value, “the interference sound exists”, and “the target sound exists”. The coherence COH (K) is a small value.

When the above behavior is organized by focusing on the presence or absence of interfering sounds, the following relationship is obtained.
When “no interference sound exists” and “target sound exists”, the coherence COH (K) becomes a large value, and the front suppression signal N (f, K) (average front suppression signal AVE_N (K)) Is a value proportional to the magnitude of the target sound component.
When “disturbance sound exists”, the coherence COH (K) has a small value, and the front suppression signal N (f, K) (average front suppression signal AVE_N (K)) has a large value.

By the way, in the case of the above behavior, if the correlation coefficient cor (K) between the front suppression signal N (f, K) (average front suppression signal AVE_N (K)) and coherence COH (K) is introduced, The same can be said.
When “no disturbing sound exists”, the correlation coefficient cor (K) is a positive value (cor (K)> 0).
When “interference sound exists”, the correlation coefficient cor (K) is a negative value (cor (K) ≦ 0).

  Therefore, the correlation and mod GI calculation unit 14 observes the sign of the correlation coefficient cor (K) between the average front suppression signal AVE_N (K) and the coherence COH (K), and the correlation coefficient cor (K) is positive. It is determined that there is no interfering sound, and it is determined that there is an interfering sound when the correlation coefficient cor (K) is negative.

  Here, the calculation method of the correlation coefficient cor (K) is not limited. For example, the correlation coefficient cor (K) can be calculated for each frame using the following equation (9). .

  In the following equation (9), cov [AVE_N (K), COH (K)] indicates the covariance between the average front suppression signal AVE_N (K) and the coherence COH (K). In the following equation (9), σAVE_N (K) represents the standard deviation of the average front suppression signal AVE_N (K), and σCOH (K) represents the standard deviation of the coherence COH (K). Furthermore, when calculating the correlation coefficient cor (K) in the following equation (9), using the result of a predetermined number i frames processed most recently for AVE_N (K) and COH (K) respectively, Standard deviation and covariance may be obtained. Specifically, in the process of obtaining the correlation coefficient cor (K) in the following (9), for example, i frames (Ki-th frame, K- (i-1) processed most recently) The standard deviations (σN (f, K) and σCOH (K)) and COH (K) and AVE_N related to the first frame,..., K−1th frame, Kth frame), Covariance (cov [AVE_N (K), COH (K)]) may be obtained. In other words, in the process of obtaining the correlation coefficient cor (K), the standard deviation and covariance in the following equation (9) may be obtained by using the i AVE_N and COH obtained most recently as samples. . The correlation coefficient cor (K) obtained in this way takes a value of −1.0 to 1.0.

  Next, the correlation and modGI calculation unit 14 uses the correlation coefficient cor (K) to calculate a feature amount representing the intensity of positive and negative fluctuations in the slope of the amplitude of the correlation coefficient cor (K).

  When background noise exists in the input signal, the behavior of the correlation coefficient cor (K) changes as follows.

  When the interference sound exists, the value of the correlation coefficient cor (K) becomes positive, and when the interference sound does not exist, the value of the correlation coefficient cor (K) becomes negative, and the macro behavior is maintained to some extent. .

  ・ Under the influence of background noise, the irregularity of the fluctuation of the amplitude of the front suppression signal (average front suppression signal AVE_N (K)) increases, whereas the coherence COH (K) decreases the dynamic range. Therefore, the irregularity of the amplitude does not change drastically. For this reason, the synchronization between the increase / decrease of the front suppression signal (average front suppression signal AVE_N (K)) and the increase / decrease of coherence COH (K) is impaired, and the increase / decrease of the correlation (correlation coefficient cor (K)). The fluctuations of the In addition, the frequency of positive and negative fluctuations in the correlation coefficient cor (K) increases.

  That is, as the influence of background noise increases, the fluctuation in the value of the correlation coefficient cor (K) increases and the frequency of positive and negative fluctuations in the value of the correlation coefficient cor (K) increases.

  As described above, when background noise exists, the frequency of increase / decrease in the value of the correlation coefficient cor (K) and the frequency of positive / negative fluctuations increase, and as the influence of background noise increases, these fluctuations (that is, the correlation) The increase / decrease in the value of the number cor_ (K) and the positive / negative fluctuation) increase. This behavior is derived only from background noise. Therefore, by observing the intensity of fluctuation of the correlation coefficient cor (K), the degree of influence of the background noise on the target sound and the fluctuation of the characteristics are estimated without being affected by the target sound and the disturbing sound. be able to.

  Therefore, in the first embodiment, the correlation and mod GI calculation unit 14 uses a feature amount called mod GI (GI: Gradient Index) in order to observe an increase or decrease in the value of the correlation coefficient cor (K) or a positive / negative fluctuation. calculate.

  Here, modGI is an index for measuring the number of changes in the inclination direction of the signal waveform and its magnitude (see Patent Document 2). modGI is defined as the power of the second-order difference of the calculation target signal, normalized with respect to the power of the calculation target signal, with respect to an arbitrary signal of the feature amount calculation target.

  In the first embodiment, the correlation and modGI calculation unit 14 calculates modGI according to the calculation method described in Patent Document 2. As an example of the calculation formula of mod GI defined as described above, the feature and the correlation and mod GI calculation unit 14 represent the intensity of fluctuation of the correlation coefficient cor (K) using the following formula (10). Calculate cor_modGI (K).

  Equation (10) represents the frequency at which the slope of the correlation coefficient cor (K) varies. Equation (10) is characterized in that the value of cor_modGI decreases as the positive / negative fluctuation of the signal slope decreases, whereas the value of cor_modGI increases as the positive / negative fluctuation of the slope increases. In other words, it can be said that the larger the value of cor_modGI, the greater the influence of background noise, and the smaller the value of cor_modGI, the smaller the influence of background noise.

  The WF unit 15 sets a value of a time constant (λ) for controlling the adaptive speed of the suppression coefficient wf_coef (f, K) based on the value of the cor_modGI (K) from the correlation and modGI calculation unit 14, and this time constant Is used to calculate the suppression coefficient wf_coef (f, K).

  The WF unit 15 multiplies the frequency domain signal X1 (f, K) of the input signal by the suppression coefficient wf_coef (f, K) to calculate the post-suppression signal Y (f, K), and the IFFT unit 16 is output.

  FIG. 4 is a block diagram showing a configuration of the WF unit 15 according to the first embodiment.

  As shown in FIG. 4, the WF unit 15 according to the first embodiment includes an input signal acquisition unit 21, a time constant control unit 23, a coefficient adaptation unit 24, a background noise suppression processing unit 25, and a post-suppression signal output unit 26. Have

  The input signal acquisition unit 21 acquires the frequency domain signal X1 (f, K) of the input signal and cor_modGI (K) from the correlation and modGI calculation unit 14.

  The time constant control unit 23 sets the value of the time constant λ that controls the adaptive speed of the suppression coefficient wf_coef (f, K) based on the value of the cor_modGI (K) from the correlation and modGI calculation unit 14.

  Here, the role of the time constant λ will be briefly described. In the WF unit 15, a suppression coefficient adaptation unit 24, which will be described later, calculates the suppression coefficient wf_coef (f, K). Prior to this, the background noise characteristic must be calculated. The background noise is estimated by, for example, Equation 1 of Patent Document 1, and a parameter (time constant) λ is involved here.

  The time constant λ takes a value of 0.0 to 1.0 and has a role of controlling how much the instantaneous input value is reflected on the background noise characteristics. The larger the value of the time constant λ, the stronger the influence of the instantaneous input, and the smaller the value of the time constant λ, the less the influence of the instantaneous input. Therefore, if the value of the time constant λ is large, the value of the suppression coefficient wf_coef (f, K) reflects the instantaneous input strongly, so that high-speed coefficient adaptation can be realized, but the influence of the instantaneous input becomes strong. The coefficient value fluctuates greatly, which may reduce the naturalness of sound quality. On the other hand, when the value of the time constant λ is small, although the adaptation speed is slow, the obtained suppression coefficient wf_coef (f, K) is not strongly influenced by the instantaneous characteristics, and the past noise characteristics are reflected on average. Therefore, it is difficult to lose the natural sound quality.

  Therefore, when the value of cor_mod (K) is larger than the threshold Θ (for example, when cor_mod (K) is greater than or equal to the threshold Θ), the time constant control unit 23 has a large influence of the background noise. Set the value to a large value. On the other hand, when the value of cor_mod (K) is smaller than the threshold value Θ (for example, when cor_mod (K) is less than the threshold value Θ), the time constant control unit 23 is less influenced by background noise and has a value of the time constant λ. Make it smaller. As a result, the coefficient adaptation according to the characteristics of the background noise can be realized without being affected by the target sound and the interference sound.

  Here, the case where the threshold value θ for judging the magnitude of the value of the time constant λ is one is illustrated, but two or more threshold values may be set, and the time constant is finely set for each section to which cor_modGI belongs. λ may be set.

  The suppression coefficient adaptation unit 24 calculates the suppression coefficient wf_coef (f, K) using the time constant λ set by the time constant control unit 23. The suppression coefficient wf_coef (f, K) can be obtained using, for example, Equation 3 in Patent Document 1.

The background noise suppression processing unit 25 uses the following equation (11) to convert the suppression coefficient wf_coef (f, K) calculated by the suppression coefficient adaptation unit 24 into the frequency domain signal X1 (f, K) of the input signal. Multiplication is performed to calculate a post-suppression signal Y (f, K).
Y (f, K) = X1 (f, K) × wf_coef (f, K) (11)

  The post-suppression signal output unit outputs the post-suppression signal Y (f, K) to the IFFT unit 16.

  The IFFT unit 16 converts a signal Y (f, K) that is a frequency domain signal into a time domain signal y (n). Note that the IFFT unit 16 may be omitted if the subsequent circuit is configured to process the frequency domain signal Y (f, K) as it is.

(A-2) Operation of the First Embodiment Next, the operation of the non-target sound suppression process in the non-target sound suppression device 1 according to the first embodiment will be described in detail with reference to the drawings.

  First, input signals s1 (n) and s2 (n) for one frame (for one processing unit) are supplied from each of the microphones m_1 and m_2 to the FFT unit 11 via an AD converter (not shown). The FFT unit 11 performs Fourier transform on the analysis frames FRAME1 (K) and FRAME2 (K) based on the input signals s1 (n) and s2 (n) for one frame, and a signal X1 (f, K) indicated in the frequency domain. , X2 (f, K). The signals X1 (f, K) and X2 (f, K) generated by the FFT unit 11 are given to the front suppression signal generation unit 12 and the coherence calculation unit 13.

  The front suppression signal generator 12 calculates a front suppression signal N (f, K) based on the signals X1 (f, K) and X2 (f, K) from the FFT unit 11. Then, the front suppression signal generation unit 12 calculates an average front suppression signal AVE_N (K) based on the front suppression signal N (f, K), and provides it to the correlation and modGI calculation unit 14.

  The coherence calculation unit 13 generates a coherence COH (K) based on the signals X1 (f, K) and X2 (f, K) from the FFT unit 11 and gives them to the correlation and modGI calculation unit 14.

  The correlation and modGI calculation unit 14 calculates a correlation coefficient cor (K), which is a feature amount indicating the relationship between the average front suppression signal AVE_N (K) and the coherence COH (K), using, for example, Expression (9). To do.

  Further, the correlation and mod GI calculation unit 14 uses the correlation coefficient cor (K), and cor_modGI (K) that is a feature amount that represents the intensity of positive and negative fluctuations in the amplitude slope of the correlation coefficient cor (K). And cor_modGI (K) is given to the WF unit 15.

  Cor_modGI (K) and the frequency domain signal X1 (f, K) of the input signal are input to the WF unit 15 from the correlation and modGI calculation unit 14.

  FIG. 5 is a flowchart showing processing in the time constant control unit 23 of the WF unit 15 according to the first embodiment.

  First, the time constant control unit 23 compares the value of cor_modGI (K) from the correlation and modGI calculation unit 14 with the threshold Θ (S101), and when the value of cor_modGI (K) is larger than the threshold Θ, the time constant λ If the value of cor_modGI (K) is less than the threshold Θ, the value of the time constant λ is set to a small value (S102).

  The time constant λ takes a value of 0.0 <λ <1.0, and as the value of the time constant λ approaches 1.0, the time constant λ is strongly influenced by a signal input at an instant. As the value of the time constant λ approaches 0.0, the influence of the signal input at the moment becomes weaker. Therefore, the value of the time constant λ based on the comparison result between the value of cor_modGI (K) and the threshold Θ can be a relative magnitude. Therefore, when the value of cor_modGI (K) is less than the threshold Θ, the value of the time constant λ is λ1, and when the value of the cor_modGI (K) is greater than or equal to the threshold Θ, the value of the time constant λ is λ2, λ1 <λ2. It is sufficient if it is a magnitude relationship.

  Then, the suppression coefficient adaptation unit 24 calculates the suppression coefficient wf_coef (f, K) using the time constant λ set by the time constant control unit 23.

  That is, the higher the time constant λ, the faster the suppression coefficient wf_coef (f, K) that strongly reflects the influence of the instantaneous input can be calculated. On the other hand, if the value of the time constant λ is small, the influence of the instantaneous input is reduced and the adaptive speed of the suppression coefficient wf_coef (f, K) is slow, but the obtained suppression coefficient wf_coef (f, K) is affected by the instantaneous characteristic. The noise characteristics of the past are reflected on average. Therefore, in this case, the natural sound quality is not easily lost.

  Further, the background noise suppression processing unit 25 uses the expression (11) to convert the suppression coefficient wf_coef (f, K) calculated by the suppression coefficient adaptation unit 24 into the frequency domain signal X1 (f, K) of the input signal. The post-suppression signal Y (f, K) is calculated by multiplication, and the post-suppression signal output unit outputs the post-suppression signal Y (f, K) to the IFFT unit 16.

  The IFFT unit 16 converts the signal Y (f, K), which is a frequency domain signal, into a time domain signal y (n) and outputs it to the subsequent audio processing apparatus 2.

(A-3) Effect of First Embodiment As described above, according to the first embodiment, the mod GI of the correlation between the front suppression signal and the coherence increases as the influence of background noise increases, and the influence is small. The time constant of the Wiener filter (WF) can be controlled based on the characteristic behavior of becoming smaller. Thereby, appropriate coefficient adaptation based on the influence of background noise becomes possible, and the accuracy of background noise suppression processing can be improved.

  Thus, by applying the present invention to a pre-processing of a communication device such as a video conference system or a mobile phone or a voice recognition function, an improvement in performance can be expected.

(B) Second Embodiment Next, a second embodiment of the non-target sound suppressing apparatus, method and program according to the present invention will be described with reference to the drawings.

  In the second embodiment, the present invention is used to subtract a front suppression signal from an input signal, for example, to subtract the interference sound coming from the surroundings, and a non-target sound suppression apparatus and method (interference sound suppression apparatus) And method).

  When subtracting the front suppression signal from the input signal, the subtraction strength is often controlled by multiplying the front suppression signal by the subtraction coefficient. If the subtraction coefficient is too large, the suppression performance will be excessive and distortion of the target speech will increase. If the subtraction coefficient is too small, the sound quality is greatly affected such that the suppression performance of the disturbing voice is insufficient. However, it is difficult to determine the presence of disturbing speech superimposed on the target speech, and it is difficult to set the subtraction coefficient to an appropriate value.

  Therefore, in the second embodiment, a non-target sound suppression device that estimates the contribution degree of the interference sound to the input signal and controls the subtraction coefficient of the frequency subtraction according to the result to suppress the interference sound without excess or deficiency. And a method (interference sound suppressing apparatus and method) are realized.

(B-1) Configuration of Second Embodiment FIG. 6 is a block diagram showing an overall configuration of a non-target sound suppressing apparatus 1A according to the second embodiment.

  The non-target sound suppressing apparatus 1A according to the second embodiment acquires input signals s1 (n) and s2 (n) from a plurality of microphones m_1 and m_2 (FIG. 1 shows two cases). Then, the contribution degree of the disturbing sound to the input signal is estimated, the subtraction coefficient of the frequency subtraction is controlled according to the result, and the post-suppression signal in which the disturbing sound is suppressed is supplied to the subsequent speech processing apparatus 2.

  As in the first embodiment, the sound processing device 2 performs predetermined sound processing using the post-suppression signal from the non-target sound suppressing device 1A.

  As illustrated in FIG. 6, the non-target sound suppression apparatus 1A includes an FFT unit 11, a front suppression signal generation unit 12, a coherence calculation unit 13, a correlation calculation unit 54, a frequency subtraction processing unit 55, and an IFFT unit 16.

  Note that the FFT unit 11, the front suppression signal generation unit 12, the coherence calculation unit 13, and the IFFT unit 16 are basically the same or corresponding components described in the first embodiment, and thus detailed description thereof is omitted. To do.

  The non-target sound suppression device 1A may be realized by installing a program (for example, a non-target sound suppression program) in a computer having a processor, a memory, and the like. In this case, the non-target sound suppression device 1A is functional. Can be shown using FIG. Note that a part or all of the non-target sound suppressing apparatus 1A may be realized by hardware.

  The correlation calculation unit 54 acquires the frontal suppression signal (average frontal suppression signal AVE_N (K)) from the frontal suppression signal generation unit 12, and the coherence COH (K) from the coherence calculation unit 13, and the average frontal suppression signal AVE_N (K). ) And the coherence COH are calculated. Further, the correlation calculation unit 54 outputs the calculated correlation coefficient cor (K) to the frequency subtraction processing unit 55. As a method for calculating the correlation coefficient cor (K), the same method as in the first embodiment can be used, and for example, the equation (9) can be used.

  The frequency subtraction processing unit 55 acquires the input signal X1 (f, K), the correlation coefficient cor (K) from the correlation calculation unit 54, and the front suppression signal N (f, K) from the front suppression signal generation unit 12. Based on the correlation coefficient cor (K), a subtraction coefficient α is set, the front suppression signal N (f, K) is multiplied by the subtraction coefficient α, and then subtracted from the input signal X1 (f, K). Then, the post-suppression signal Y (f, K) is obtained.

  FIG. 7 is a block diagram illustrating a configuration of the frequency subtraction processing unit 55 according to the second embodiment.

  As illustrated in FIG. 7, the frequency subtraction processing unit 55 includes an input signal acquisition unit 31, a subtraction coefficient control unit 32, a subtraction unit 33, and a post-subtraction processing signal output unit 34.

  The input signal acquisition unit 31 acquires the input signal X1 (f, K), the correlation coefficient cor (K) from the correlation calculation unit 54, and the front suppression signal N (f, K) from the front suppression signal generation unit 12. Is.

  The subtraction coefficient control unit 32 sets the subtraction coefficient α based on the correlation coefficient cor (K).

  Here, the principle of estimating the contribution of the disturbing sound (here, the disturbing sound) will be described. First, it is assumed that the target sound comes from the front of the microphones m_1 and m_2, and the disturbing sound comes from the lateral direction (right direction, left direction) of the microphones m_1 and m_2.

  At this time, since the front suppression signal N (f, K) captures the signal component coming from the front when “no disturbing sound exists” and “the target sound exists”, the magnitude of the target sound component is large. It has a signal value proportional to the height. However, as shown in FIG. 2, the sound collection level in the front direction is smaller than that in the horizontal direction, and is smaller than the case where “disturbance sound exists”.

  The coherence COH is a feature quantity that has a deep relationship with the arrival direction of the input signal. Therefore, the value is large when “no disturbing sound exists” and “only the target sound exists”, and small when “disturbing sound exists”.

  The above behavior can be summarized as follows, focusing on the presence or absence of interfering sounds.

  When “no disturbing sound exists” and “only the target sound exists”, the coherence COH is a large value, and the front suppression signal is a value proportional to the size of the target sound component.

  When “interfering sound is present”, the coherence COH is a small value and the front suppression signal is a large value.

  This behavior is as follows when the correlation coefficient cor (K) between the front suppression signal N (f, K) and the coherence COH is introduced.

  In the case of “no disturbing sound”, the correlation coefficient cor (K) is a positive value.

  In the case of “no disturbing voice”, the correlation coefficient cor (K) is a negative value.

  By the way, the subtraction coefficient α is preferably a smaller value as the influence of the disturbing sound is smaller and a larger value as the influence of the disturbing sound is larger from the viewpoint of reducing the excess or deficiency of the disturbing sound (described later (12)). See formula).

  As described above, the sign varies depending on the presence or absence of the interfering sound. If the correlation coefficient cor (K) is positive, the subtraction coefficient α is decreased. If the correlation coefficient (K) is negative, the subtraction coefficient α is increased. By such processing, the subtraction coefficient can be controlled in accordance with the influence level of the disturbing sound.

  Therefore, in the second embodiment, the subtraction coefficient control unit 32 uses the subtraction used for the frequency subtraction process based on the behavior peculiar to the correlation coefficient cor (K) between the front suppression signal N (f, K) and the coherence COH. Control the coefficient.

  More specifically, the subtraction coefficient control unit 32 sets a large value for the subtraction coefficient α in order to enhance the suppression effect when the disturbing sound is present, and exhibits the suppression effect when there is no disturbing sound. In order to weaken, a small value is set for the subtraction coefficient α.

  The subtraction coefficient control unit 32 includes, for example, a subtraction coefficient storage unit (not shown) that records the correspondence between the correlation coefficient value and the set value of the subtraction coefficient α, and refers to this subtraction coefficient storage unit. Thus, the subtraction coefficient α may be set.

The subtraction unit 33 uses the subtraction coefficient α obtained from the subtraction coefficient control unit 32 to perform a subtraction process as shown in equation (12).
Y (f, K) = X1 (f, K) −α × N (f, K) (12)

  The post-subtraction signal output unit 34 outputs the post-suppression signal (subtraction signal) Y (f, K) calculated by the subtraction unit 33 to the IFFT unit 16.

(B-2) Operation of Second Embodiment Next, the operation of the non-target sound suppression process in the non-target sound suppression device 1A according to the second embodiment will be described in detail with reference to the drawings.

  Input signals s1 (n) and s2 (n) for one frame (for one processing unit) are supplied to the FFT unit 11 from each of the microphones m_1 and m_2 via an AD converter (not shown). The FFT unit 11 performs Fourier transform on the analysis frames FRAME1 (K) and FRAME2 (K) based on the input signals s1 (n) and s2 (n) for one frame, and a signal X1 (f, K) indicated in the frequency domain. , X2 (f, K). The signals X1 (f, K) and X2 (f, K) generated by the FFT unit 11 are given to the front suppression signal generation unit 12 and the coherence calculation unit 13.

  The front suppression signal generator 12 calculates a front suppression signal N (f, K) based on the signals X1 (f, K) and X2 (f, K) from the FFT unit 11. Then, the front suppression signal generation unit 12 calculates an average front suppression signal AVE_N (K) based on the front suppression signal N (f, K), and gives it to the correlation calculation unit 54.

  The coherence calculation unit 13 generates coherence COH (K) based on the signals X1 (f, K) and X2 (f, K) from the FFT unit 11 and gives them to the correlation calculation unit 54.

  The correlation calculation unit 54 calculates a correlation coefficient cor (K), which is a feature amount indicating the relationship between the average front suppression signal AVE_N (K) and the coherence COH (K), using, for example, equation (9).

  The frequency subtraction processing unit 55 receives the input signal X1 (f, K), the correlation coefficient cor (K) from the correlation calculation unit 54, and the front suppression signal N (f, K) from the front suppression signal generation unit 12. Is done.

  FIG. 8 is a flowchart showing processing in the subtraction coefficient control unit 32 of the frequency subtraction processing unit 55 according to the second embodiment.

  First, the subtraction coefficient control unit 32 determines whether or not the value of the correlation coefficient cor (K) from the correlation calculation unit 54 is negative (S201). When the value of the correlation coefficient cor (K) is negative (that is, when disturbing voice is present), a large value is set for the subtraction coefficient α in order to enhance the suppression effect (S202). On the other hand, when the value of the correlation coefficient cor (K) is not negative (that is, when no disturbing sound exists), a small value is set for the subtraction coefficient α in order to weaken the suppression effect.

  Then, the subtraction unit 33 uses the subtraction coefficient α obtained by the subtraction coefficient control unit 32 to obtain a post-subtraction signal Y (f, K) by the equation (12), and the subtraction signal output unit 34 Then, the signal after suppression processing (the signal after subtraction processing) Y (f, K) is output to the IFFT unit 16.

  The IFFT unit 16 converts the signal Y (f, K), which is a frequency domain signal, into a time domain signal y (n) and outputs it to the subsequent audio processing apparatus 2.

(B-3) Effect of the Second Embodiment As described above, according to the second embodiment, when the disturbing speech exists, the correlation coefficient between the front suppression signal and the coherence is negative, and the disturbing speech is Based on the characteristic behavior of being positive when it does not exist, the presence of interfering speech superimposed on the target speech is detected, and this result is used to control the subtraction coefficient used in the frequency subtraction process. The accuracy of the voice suppression process can be increased.

  Thus, by applying the present invention to a pre-processing of a communication device such as a video conference system or a mobile phone or a voice recognition function, an improvement in performance can be expected.

(C) Other Embodiments Although various modified embodiments are mentioned in the first and second embodiments described above, the present invention can also be applied to the following modified embodiments.

  (C-1) In the first or second embodiment described above, the suppression coefficient or the subtraction coefficient may be calculated for each frequency bin. In this case, the correlation coefficient can also be realized by calculating for each frequency bin.

  (C-2) In the second embodiment, the presence / absence of the interfering sound can be determined by paying attention to the positive / negative of the correlation coefficient. I understand. The specific relationship between the correlation coefficient and the influence of the interfering sound is as follows. If the correlation coefficient is negative and the absolute value is small, the influence of the interfering sound is small. It ’s big. Therefore, an arbitrary function (for example, a quadratic function) is prepared so that the output value is small if the input value is small and the output value is large if the input value is large, and the absolute value of the correlation coefficient is input to this. By using the obtained value as the subtraction coefficient, it is possible to set a subtraction coefficient corresponding to the influence level of the interference sound (the magnitude of the absolute value of the correlation).

  DESCRIPTION OF SYMBOLS 1 and 1A ... Non-target sound suppression apparatus, 11 ... FFT part, 12 ... Front suppression signal generation part, 13 ... Coherence calculation part, 14 ... Correlation and modGI calculation part, 15 ... WF (Wiener filter) part, 54 ... Correlation calculation , 55... Frequency subtraction processing unit, 16... IFFT unit.

Claims (7)

Front-side suppression signal generation that generates a front-side suppression signal having a blind spot in front based on the difference between the plurality of frequency-domain input signals obtained by converting each input signal from each of the plurality of microphones from the time domain to the frequency domain And
A coherence calculator that calculates coherence based on signals obtained from the plurality of input signals;
A feature amount calculating unit that calculates a feature amount indicating a relationship between the front suppression signal and the coherence;
Using a feature amount indicating the relationship between the front suppression signal and the coherence, a coefficient related to suppression of the non-target sound included in the input signal is set, and the non-purpose included in the input signal using the coefficient A non-target sound suppression apparatus comprising: a non-target sound suppression processing unit that obtains a signal after suppression processing that suppresses sound. The feature amount calculation unit calculates a feature amount that represents the intensity of positive and negative fluctuations in the slope of the correlation amplitude indicating the relationship between the front suppression signal and the coherence,
The non-target sound suppression processing unit sets a variable used for background noise suppression processing using the feature amount representing the intensity of positive and negative fluctuations in the slope of the correlation amplitude, and the suppression obtained using the variable The non-target sound suppression apparatus according to claim 1, wherein a signal after suppression processing is obtained by multiplying the input signal by a coefficient. The feature amount calculation unit calculates the feature amount by normalizing the power of the second-order difference of the correlation indicating the relationship between the front suppression signal and the coherence with the power of the correlation,
The non-target sound suppression apparatus according to claim 2, wherein the non-target sound suppression processing unit sets a variable used for the background noise suppression processing according to a comparison result between the feature amount and a threshold value. The feature amount calculating unit calculates a feature amount indicating a correlation indicating a relationship between the front suppression signal and the coherence;
The non-target sound suppression processing unit sets a subtraction coefficient using the feature amount representing the correlation, subtracts a product of the front suppression signal and the subtraction coefficient from the input signal, and outputs a signal after suppression processing. The non-target sound suppression device according to claim 1, wherein the non-target sound suppression device is obtained. The non-target sound suppression processing unit sets the subtraction coefficient according to the sign of the feature value representing the correlation, and subtracts the product of the front suppression signal and the subtraction coefficient from the input signal, The non-target sound suppression apparatus according to claim 5, wherein a post-suppression signal is obtained. A front suppression signal having a blind spot in front based on a difference between a plurality of frequency domain input signals obtained by converting each input signal from each of a plurality of microphones from a time domain to a frequency domain by a front suppression signal generation unit. Produces
A coherence calculator calculates coherence based on signals obtained from the plurality of input signals;
A feature amount calculation unit calculates a feature amount indicating a relationship between the front suppression signal and the coherence;
The non-target sound suppression processing unit sets a coefficient related to suppression of the non-target sound included in the input signal using the feature amount indicating the relationship between the front suppression signal and the coherence, and uses the coefficient A non-target sound suppression method, comprising: obtaining a post-suppression signal that suppresses a non-target sound included in the input signal. Computer
Front-side suppression signal generation that generates a front-side suppression signal having a blind spot in front based on the difference between the plurality of frequency-domain input signals obtained by converting each input signal from each of the plurality of microphones from the time domain to the frequency domain And
A coherence calculator that calculates coherence based on signals obtained from the plurality of input signals;
A feature amount calculating unit that calculates a feature amount indicating a relationship between the front suppression signal and the coherence;
Using a feature amount indicating the relationship between the front suppression signal and the coherence, a coefficient related to suppression of the non-target sound included in the input signal is set, and the non-purpose included in the input signal using the coefficient A non-target sound suppression program that functions as a non-target sound suppression processing unit that obtains a signal after suppression processing that suppresses sound.

Images