JP6517124B2 - Noise suppression device, noise suppression method, and program - Google Patents

Description

  The present invention relates to a technique for suppressing noise from an acoustic signal picked up by a multi-channel microphone array to emphasize a speaker's voice.

  Among the techniques for performing noise suppression using a multi-channel microphone array, a directional microphone array based on a delay-and-sum array is often used as a beamforming technique for which direction control is easy (for example, see Non-Patent Document 1) . In the method described in Non-Patent Document 1, the delay between microphones when sound arriving from a specific direction reaches a multi-channel microphone is calculated, and the sound reception signal of the microphone with a small delay Delay according to the incoming signal. As a result, the signal in the specific direction is in-phase with all the sound receiving signals, and the output gain is maximized when the sum of the signals is output.

  As multi-channel noise suppression using a directional microphone array, there is a method of applying a filter to the subsequent stage of beam forming (see, for example, Patent Document 1). In the method described in Patent Document 1, the amount of signal is estimated from the signal collected for each direction, and a Wiener filter is designed using the estimated amount of signal.

Patent No. 4724054 gazette

Toshiro Oga, Yutaka Kanada, Yoshio Yamazaki, "Sound System and Digital Processing", The Institute of Electronics, Information and Communication Engineers, 1995

  However, in the situation where many sound sources are densely present, it is difficult to separate the sound sources based only on spatial information because the signals from the respective sound sources interfere with each other. Therefore, in such a situation, there is a problem that the performance of noise suppression is degraded in the conventional beamforming.

  An object of the present invention is to provide a noise suppression technique capable of modeling a signal arriving from a specific direction with high accuracy and enhancing noise suppression performance in view of such a problem.

  In order to solve the above problems, the noise suppression device according to the present invention represents each output probability of at least one of a clean speech signal and a noise signal and a silent signal by a mixed normal distribution containing a plurality of normal distributions. A model parameter storage unit for storing parameters of the signal component model, and a plurality of frequencies obtained by dividing each of a plurality of angle region signals picked up by emphasizing sounds coming from a plurality of angle regions in different directions into a plurality of band components It is expressed by a mixed normal distribution including a plurality of normal distributions using a specific direction selection unit that selects a specific direction frequency domain signal belonging to an angular domain signal of an angular domain in a desired direction from the domain signal and parameters of the signal component model. Specific direction signal model, and calculate the posterior probability of the specific direction signal model when the specific direction frequency domain signal is observed. Direction signal modeling unit, gain coefficient calculation unit for weighted addition of gain coefficients for each frequency band calculated using parameters of signal component model with posterior probability of specific direction signal model, and gain coefficients in specific direction frequency domain And a multiplication unit for multiplying the signal amount of each corresponding frequency band of the signal.

  According to the noise suppression technique of the present invention, it is possible to model a signal coming from a specific direction with high accuracy by using together the signal amount of each angle region and the signal amount based on the statistical model. By using this model, the Wiener filter can be designed with high accuracy, and the noise suppression performance is improved.

FIG. 1 is a diagram illustrating a functional configuration of the noise suppression device of the first embodiment. FIG. 2 is a diagram illustrating the processing procedure of the noise suppression method of the first embodiment. FIG. 3 is a diagram illustrating the functional configuration of the model parameter storage unit. FIG. 4 is a diagram for explaining the directivity characteristic of the beam former. FIG. 5 is a diagram for explaining an example of division of the directivity direction area set in the beam former. FIG. 6 is a diagram illustrating the functional configuration of the beam former. FIG. 7 is a diagram illustrating a functional configuration of the specific direction selection unit. FIG. 8 is a diagram illustrating a functional configuration of the signal amount estimating unit. FIG. 9 is a diagram for explaining an example of the directivity characteristic of the beam former. FIG. 10 is a diagram illustrating a functional configuration of the specific direction signal modeling unit of the first embodiment. FIG. 11 is a diagram illustrating a functional configuration of the gain coefficient calculation unit. FIG. 12 is a diagram illustrating a functional configuration of the noise suppression device of the second embodiment. FIG. 13 is a diagram illustrating a functional configuration of the specific direction signal modeling unit of the second embodiment. FIG. 14 is a diagram illustrating a functional configuration of the noise suppression device of the third embodiment. FIG. 15 is a diagram illustrating a functional configuration of the specific direction signal modeling unit of the third embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In the drawings, components having the same functions are denoted by the same reference numerals and redundant description will be omitted.

First Embodiment
In the prior art, when multiple sound sources are densely present, the sound from each sound source interferes with each other, so that the estimation error of the power spectral density (PSD) estimated from the collected sound signal becomes large. Therefore, in the first embodiment, a clean speech model for each state is learned in advance, and the estimation error of the speech PSD included in the collected sound signal is corrected using the clean speech model. By designing a Wiener filter using the corrected voice PSD, it is possible to suppress noise with higher accuracy than ever before.

  In each embodiment of the present invention, an audio signal (clean audio signal and silent signal) and a noise signal are defined as follows. Even when recording is performed in a soundproof room or the like in which no noise is present, very small and white noise is observed in the recorded signal. A signal observed in such an environment is defined as a "silence signal". A silence signal can also be said to be a type of noise, but this noise is noise that is generated due to electrical factors such as an electrical circuit of a recording equipment or a transfer system. On the other hand, the running noise and the wind noise of a car are noises generated by acoustic factors which are observed when sound waves are transmitted through the atmosphere. Therefore, noise due to electrical factors and noise due to acoustic factors are distinguished, and only the latter is defined as "noise signal". Also, when speech is performed in an environment where a silence signal is observed, the speech signal is observed in a form superimposed on the silence signal. This superimposed signal is defined as "clean speech signal". Then, in an environment where no noise signal is present, a clean speech signal is observed between continuous silence signals. The silence signal and the clean speech signal are generically defined as "voice signal".

  As shown in FIG. 1, the noise suppression apparatus according to the first embodiment includes a model parameter storage unit 10, a microphone array 11 consisting of M (22) microphones 11-1 to 11-M, and Q (≧ 2) Beamformer units 12-1 to 12-Q, Q frequency domain conversion units 13-1 to 13-Q, specific direction selection unit 14, signal amount estimation unit 15, vector element extraction unit 16, specific direction signal model And a gain coefficient calculation unit 18, a multiplication unit 19, and an inverse frequency domain conversion unit 20. The noise suppression method of the first embodiment is realized by the processing of each step described later by this noise suppression device.

  The noise suppression device is, for example, a special program configured by reading a special program into a known or dedicated computer having a central processing unit (CPU: Central Processing Unit), a main memory (RAM: Random Access Memory), etc. It is an apparatus. The noise suppression device executes each process, for example, under the control of the central processing unit. The data input to the noise suppressor and the data obtained by each process are stored, for example, in the main storage device, and the data stored in the main storage device is read out as needed and used for other processing. Ru. In addition, at least a part of each processing unit of the noise suppression device may be configured by hardware such as an integrated circuit. Each storage unit of the noise suppression device is, for example, a main storage device such as a random access memory (RAM), an auxiliary storage device configured by a semiconductor memory device such as a hard disk, an optical disk, or a flash memory, or a relational database. And middleware such as a key value store.

  The processing procedure of the noise suppression method of the first embodiment will be described with reference to FIG.

(Model parameter storage unit)
As shown in FIG. 3, the model parameter storage unit 10 includes a silent GMM storage unit 10A and a clean voice GMM storage unit 10B. In the silent GMM storage unit 10A and the clean voice GMM storage unit 10B, parameters of a probability model of a silent signal and a clean voice signal prepared in advance are stored, respectively. In this embodiment, a mixed normal distribution model (GMM: Gaussian Mixture Model) containing a plurality of normal distributions is used as a probability model. The estimation accuracy improves as the number of normal distributions included in the mixture normal distribution model increases, but the processing speed decreases. Therefore, the number of normal distributions included in the mixed normal distribution model is desirably a value between 2 and 512 effectively, and approximately 32 is most desirable. Further, each normal distribution is configured using the mixing weight w ^s _{j, k} , the average μ ^s _{j, k} (ω), and the variance σ ^s _{j, k} ² (ω) as parameters. Here, j is the type of GMM (j = 0: silent GMM, j = 1: clean voice GMM), k is the number of each normal distribution, and ω is the frequency bin number. For example, the component of each frequency bin of signal power is used for the feature quantity of GMM. In addition, since many well-known techniques and well-known techniques exist about the structure method of GMM, detailed description is abbreviate | omitted (for example, refer the following reference 1 grade | etc.,).
[Reference 1] Nakagawa Seiichi, "Speech recognition based on probabilistic models", The Institute of Electronics, Information and Communication Engineers

That is, the model parameter storage unit 10 is a model parameter of a probability model (hereinafter also referred to as a speech model parameter) in which each output probability of the clean speech signal and the silence signal is represented by a mixed normal distribution containing a plurality of normal distributions. Is stored. The speech model parameter ^{S S} is configured by equation (1).

(Microphone array)
In step S11, the microphone array 11 picks up M signals x _m (n) (m = 1,..., M) using the M microphones 11-1 to 11-M. Here, n represents the sample number of the discrete time signal. The signals x _m (n) are input to the beam formers 12-1 to 12-Q, respectively.

(Beam Former)
In step S12, the beam former unit 12-1 to 12-Q is, for example, the directivity of the beam BM as shown in FIG. 4, one of Q direction region theta ₁ through? _Q previously given in Figure 5 A process is performed to emphasize and collect the sound emitted in the corresponding direction area toward the heel, and the result is output. The output signals y ₁ (n), y ₂ (n),..., Y _Q (n) of the beam formers 12-1 to 12-Q are input to the frequency domain transformers 13-1 to 13-Q, respectively. Ru.

FIG. 6 shows the configuration of one of the beam formers 12-1 to 12-Q, for example, the beam former 12-q. Similar processing is performed in all beamformers. The input signals x ₁ (n) to x _M (n) are input to the filter processing units FC-1 to FC-M, respectively. The filter processing units FC-1 to FC-M output a signal x ' _qm (n) obtained by substituting the filter coefficient W _qm (n) given in advance into the convolution operation shown in the equation (2).

Output signals of the filter processing units FC-1 to FC-M are input to the addition unit ADD. The adder ADD adds the input signals as shown in equation (3) to obtain the output signal y _q (n) (q = 1,..., Q) of the beam former.

The filter coefficients W _qm (n) are emitted in the Q characteristic direction Θ _Q given in advance of the directivity characteristics D _q (ω, θ) of each of the beam formers 12-1 to 12-Q shown in FIG. It is configured to emphasize sound to be received and receive sound, and to suppress sound emitted in other directions.

(Frequency domain converter)
In step S13, the frequency domain conversion units 13-1 to 13-Q input signals y ₁ (n), y ₂ (n),..., Y _Q (n) have short time lengths (eg, sampling frequency 16,000 Hz) Is decomposed into frames of about 256 samples, and discrete Fourier transform is performed in each frame, for example, Ω frequency components obtained are output signals Y ₁ (ω, l), Y ₂ (ω, l), ... , Y _Q (ω, l). Here, ω represents a frequency bin number, and l represents a frame number. The output signals Y ₁ (ω, l), Y ₂ (ω, l),..., Y _Q (ω, l) of the frequency domain conversion units 13-1 to 13 -Q are the specific direction selection unit 14 and the signal amount estimation The data is input to the unit 15 respectively.

(Specific direction selection unit)
In step S14, the specific direction selection unit 14 selects and outputs a signal corresponding to the output of the beam former which directed the directional beam to the direction region to be emphasized from the output signal of the input frequency domain converter. . The output signal Y _S (ω, l) of the specific direction selection unit 14 is input to the specific direction signal modeling unit 17 and the multiplication unit 19.

FIG. 7 shows the configuration of the specific direction selection unit 14. In the specific direction selection unit 14, the frequency domain signals Y ₁ (ω, l) and Y ₂ (ω which are input from the beamformer units 12-1 to 12 -Q through the frequency domain conversion units 13-1 to 13 -Q. , l), ..., Y _Q (ω, l), which corresponds to the qs direction region to be emphasized, is selected and output as Y _s (ω, l).

(Signal amount estimation unit)
In step S15, the signal amount estimation unit 15 obtains the power component of the sum of sound signals emitted from the sound source in each direction region Θ _{1 to} Θ _Q from the input output signal of the frequency domain conversion unit, The estimated signal power vector X _est (ω, l) summarized in FIG. The estimated signal power vector X _est (ω, l) output from the signal amount estimation unit 15 is input to the vector element extraction unit 16.

FIG. 8 shows the configuration of the signal amount estimation unit 15. The frequency domain signals Y ₁ (ω, l), Y ₂ (ω, l),..., Y _Q (ω, l) input to the signal amount estimation unit 15 are power operation units PW-1 to PW-Q, respectively. Is input to The power calculation units PW-1 to PW-Q have power values | Y ₁ (ω, l) | ² , | Y ₂ (ω, l) | ² , ..., | Y _Q (ω, l) | Calculate ² and output the result. The power values | Y ₁ (ω, l) | ² , | Y ₂ (ω, l) | ² ,..., | Y _Q (ω, l) | ² are input to the area aggregation unit 15A. The area aggregation unit 15A obtains the average of the power values of the signals emitted from the set S of the predetermined areas to be picked up and the power values of the signals emitted from the set N of the areas to be suppressed, Output the aggregated power vector Y (ω, l).

Here, N _S indicates the number of regions included in the set S, and N _N indicates the number of regions included in the set N. Also, all direction areas (1 to Q) are determined in advance to belong to the set S or the set N. For example, when Q = 4, sets S and N may be defined as S = {1, 2} and N = {3, 4}, respectively.

The aggregated power vector Y (ω, l) is input to the multiplication unit 15B. The power estimation matrix T ⁻¹ (ω), which is the other input of the multiplier 15B, is an output signal of the inverse matrix calculator 15C. The aggregation gain matrix T (ω) defined by the equation (6) is input to the inverse matrix operation unit 15C, and the inverse matrix T ⁻¹ (ω) is output.

  Each element of the aggregation gain matrix T is a parameter obtained from an average value of directivity characteristics for each direction area of each of the beam formers 12-1 to 12-Q as shown in FIG. 9, for example, equation (7) Use the average value for the direction of directivity as shown in.

ε _pq (ω) is an average value of the directivity characteristics for the q-th direction region of the beam former 12-p. The directivity characteristic can be _{obtained from} the filter coefficient W _qm (n), for example, using a known technique described in the above-mentioned Non-Patent Document 1.

The multiplying unit 15B performs multiplication of the integrated power vector Y (ω, l) and the power estimation matrix T ⁻¹ (ω), which are input as shown in equation (8), for each frequency component, and estimates the estimated signal power vector X _est ( Output ω, l).

In the estimated signal power vector X _est (ω, l), as shown in equation (9), the first component is the sound collection region signal estimated power | S _bf (ω, l) | ² and the second component is the suppression region signal estimation power _{| N bf (ω, l)} | 2 to become.

  As described above, the signal amount estimating unit 15 estimates the power (signal amount) of the signal by aggregating the direction regions in the sound collection region and the suppression region.

The processing of the signal amount estimation unit is not limited to the method described above, and a known signal amount estimation technique can be used (see, for example, Patent Document 1 and Reference Document 2 below).
[Reference 2] Japanese Patent No. 4856662

(Vector element extraction unit)
In step S16, the vector element extraction unit 16 extracts the suppression region signal estimated power | N _bf (ω, l) | ² from the input estimated signal power vector X _est (ω, l) and outputs it. An output signal | N _bf (ω, l) | ² of the vector element extraction unit 16 is input to the specific direction signal modeling unit 17 and the gain coefficient calculation unit 18, respectively.

(Specific Direction Signal Modeling Unit)
In step S17, the specific direction signal modeling unit 17 uses the output signal | N _bf (ω, l) | ² of the vector element extraction unit 16 and the speech model parameter ^{S S} stored in the model parameter storage unit 10. A specific direction signal model for correcting the signal amount estimated by the signal amount estimation unit 15 is generated. Also, using the specific direction signal model, the posterior probability α _j (l) of each state j and the posterior probability β _{j, k of} each distribution k when the specific direction signal Y _S (ω, l) is observed Calculate l) The posterior probabilities α _j (l) and β _{j, k} (l) output by the specific direction signal modeling unit 17 represent how often each state and distribution of the specific direction signal model is observed in the specific direction signal observed It means the degree of lightness. The posterior probabilities α _j (l) and β _{j, k} (l) output from the specific direction signal modeling unit 17 are input to the gain coefficient calculation unit 18.

FIG. 10 shows the configuration of the specific direction signal modeling unit 17. The state probability storage unit 17C stores the probability a _j that the voice is in the state j. The suppression region signal estimation power | N _bf (ω, l) | ² and the speech model parameter ^{S S} input to the specific direction signal modeling unit 17 are input to the model parameter generation unit 17A.

The model parameter generation unit 17A calculates the parameters of the specific direction signal model as follows. Since the specific direction signal is expressed as the sum of the clean speech and the noise, the average μ ^Ys _{j, k} (ω, l) of the specific direction signal model is calculated by equation (10).

Further, the variance σ ^Ys _{j, k} ² (ω) of the specific direction signal model and the mixing weight w ^Ys _{j, k} are respectively given by equations (11) and (12).

Then, the specific direction signal model parameter Y ^YS defined by the equation (13) is output. The specific direction signal model parameter Y ^YS is input to the probability calculation unit 17B.

The specific direction signal Y _S (ω, l), the specific direction signal model parameter Y ^YS , and the state probability a _j are input to the probability calculation unit 17B. Probability calculation unit 17B is likelihood p in a particular direction signal relating to the state j from a particular direction signal model parameters [Phi ^YS | likelihood p of (Y _S j) to the specific direction signal on the distribution _{k (Y S | j, k} ) and Calculate Then, equation (14) is calculated using the likelihood p (Y _s | j), p (Y _s | j, k) of the specific direction signal and the specific direction signal model parameter 事後^YS, and the posterior probability of the distribution k Calculate β _{j, k} (l).

Also, the equation (15) is calculated using the likelihood p (Y _s | j), p (Y _s | j, k) of the specific direction signal and the state probability a _j, and the posterior probability α _j (l Calculate).

α _j (l) may be expressed by equations (16) and (17).

Here, | S _bf (ω, l) | ² is an element left unextracted by the vector element extraction unit, and is a speech PSD. Also,

は、それぞれ|S_bf(ω,l)|², |N_bf(ω,l)|²の周波数平均を示す。

_Denotes the frequency average of | S _bf (ω, l) | ² and | N _bf (ω, l) | ^{2 respectively} .

(Gain coefficient calculation unit)
In step S18, the gain coefficient calculation unit 18, the model parameter storage unit 10 reads the speech model parameters [Phi ^S, the output signal of the vector element extraction unit _{16 | N bf (ω, l} ) | 2 and a specific direction signal modeling unit The gain coefficient R (ω, l) of the Wiener filter is calculated using the posterior probabilities α _j (l) and β _{j, k} (l) received from the number 17. The gain coefficient R (ω, l) output from the gain coefficient calculator 18 is input to the multiplier 19.

FIG. 11 shows the configuration of the gain coefficient calculation unit 18. The suppression region signal estimated power | N _bf (ω, l) | ² and the speech model parameter ^{S S} input to the gain coefficient calculation unit 18 are input to the SN ratio estimation unit 18A.

  The gain factor of the basic Wiener filter is calculated by equation (18).

In this embodiment, in order to improve the accuracy of the speech PSD used for gain coefficient calculation, the SN ratio estimation unit 18A replaces the speech PSD with the speech model parameter Φ ^S with respect to the Wiener filter defined by equation (18). The gain coefficient R _{j, k} (ω, l) is calculated by the equation (19).

The weighted addition unit 18B receives the output R _{j, k} (ω, l) of the SN ratio estimation unit 18A and the outputs α _j (l) and β _{j, k} (l) of the specific direction signal modeling unit 17 (20) to calculate the gain coefficient R (ω, j) by weighted addition of the gain coefficient R _{j, k} (ω, l) with the posterior probability α _j (l), β _{j, k} (l) Do.

(Multiplication unit)
In step S19, the multiplication unit 19 outputs the result of multiplying the input gain coefficient R (ω, l) and the output signal Y _S (ω, l) of the specific direction selection unit 14 for each component of the same frequency. . The output signal Y _SR (ω, l) of the multiplication unit 19 is input to the inverse frequency domain conversion unit 20.

(Inverse frequency domain converter)
In step S20, the inverse frequency domain transformation unit 20 performs inverse discrete Fourier transform on the input output signal Y _SR (ω, l) of the multiplication unit 19 to obtain the signal y (n) restored to the time signal. Output. The output signal y (n) is a signal picked up by emphasizing the desired sound by the noise suppressor.

  In the first embodiment, the noise suppression device includes the microphone array and the Q beam formers, and the configuration for emphasizing and collecting the sound emitted in each of the Q direction regions has been described. It is not limited to such a configuration as long as it is possible to emphasize and pick up the sound emitted by the. For example, even if the noise suppressor includes Q directional microphones instead of the microphone array and the Q beam formers, the sound emitted in each direction region may be emphasized and picked up. Good. In addition, even in the configuration using directional microphones, it is possible to emphasize and collect the sound emitted in each direction area with higher accuracy by using the beam former together. Such a configuration can be similarly applied to the embodiments described later.

Second Embodiment
In the first embodiment, the estimation error of the voice PSD contained in the collected signal is corrected using a clean voice model. If the sound collection environment is limited and the nature of noise can be assumed, a noise signal similar to the expected noise can be prepared and the noise model for each state can be learned in advance. In the second embodiment, this noise model is used to correct the estimation error of the noise PSD contained in the collected signal.

  Hereinafter, the configuration of the noise suppression apparatus of the second embodiment and the processing procedure of the noise suppression method will be described focusing on differences from the first embodiment.

In the noise suppression apparatus of the second embodiment, as shown in FIG. 12, the noise model parameter ^{N N} is stored in the model parameter storage unit 10, and the output of the vector element extraction unit 16 is the sound collection region signal estimation power | S. _{bf (ω,} l) | ² to become.

(Model parameter storage unit)
In the model parameter storage unit of the second embodiment, model parameters (hereinafter referred to as noise model parameters) of a probability model in which each output probability of a noise signal and a silence signal is represented by a mixed normal distribution containing a plurality of normal distributions, respectively. ) Is stored. The noise model parameter ^{N N} is expressed by equation (21) using μ ^N _{j, k} (ω) as the average of the respective normal distributions, σ ^N _{j, k} ² (ω) as the variance, and w ^N _{j, k} as the mixing weight Configured

(Vector element extraction unit)
The vector element extraction unit 16 of the second embodiment extracts the sound collection area signal estimated power | S _bf (ω, l) | ² from the input estimated signal power vector X _est (ω, l) and outputs it. The output signal | S _bf (ω, l) | ² of the vector element extraction unit 16 is input to the specific direction signal modeling unit 17 and the gain coefficient calculation unit 18, respectively.

(Specific Direction Signal Modeling Unit)
The specific direction signal modeling unit 17 of the second embodiment, the vector output signal of the element extracting section 16 | using a ² and the model parameter noise model parameters stored in the storage unit ^{_{10 Φ N | S bf (ω}} , l) To generate a specific direction signal model. Specifically, the model parameter generation unit 17A calculates the parameter Y ^YS of the specific direction signal model by Equations (22) to (24).

(Gain coefficient calculation unit)
Second embodiment of a gain factor calculating section 18, the model parameter storage unit 10 reads the noise model parameter [Phi ^N, the output signal of the vector element extraction unit _{16 | S bf (ω, l} ) | 2 and a specific direction signal modeling Using the posterior probabilities α _j (l) and β _{j, k} (l) received from the unit 17, the gain coefficient R (ω, l) of the Wiener filter is calculated. Specifically, the SN ratio estimation unit 18A replaces the noise PSD with the noise model parameter ^{N N} with respect to the Wiener filter defined by the above equation (18), and the gain coefficient R _{j, k} by the equation (25) Calculate (ω, l).

Third Embodiment
In the first embodiment, the clean speech model is used to correct the estimation error of the speech PSD. On the other hand, in the second embodiment, the estimation error of the noise PSD is corrected using a noise model. In the third embodiment, both the clean speech model and the noise model are learned in advance, and the estimation error is corrected for both the speech PSD and the noise PSD.

  Hereinafter, the configuration of the noise suppression device of the third embodiment and the processing procedure of the noise suppression method will be described focusing on differences from the first embodiment.

The noise suppression device according to the third embodiment does not include the signal amount estimation unit 15 and the vector element extraction unit 16 as shown in FIG. 14, and the model parameter storage unit 10 has a speech model parameter ^{S S} and a noise model parameter. Φ ^N is stored.

(Specific Direction Signal Modeling Unit)
The specific direction signal modeling unit 17 of the third embodiment generates a specific direction signal model using the speech model parameter ^{S S} and the noise model parameter ^{N N} stored in the model parameter storage unit 10. Specifically, the model parameter generation unit 17A calculates the parameter Y ^YS of the specific direction signal model by Equations (26) to (28).

Here, j _S and j _N indicate the state numbers of the speech model and the noise model, and k _S and k _N indicate the distribution numbers of the speech model and the noise model, respectively.

(Gain coefficient calculation unit)
The gain coefficient calculation unit 18 of the third embodiment reads the speech model parameter ^{S S} and the noise model parameter ^{N N} from the model parameter storage unit 10, and receives the posterior probability α _j (l) received from the specific direction signal modeling unit 17. The gain coefficient R (ω, l) is calculated using β _{j, k} (l). Specifically, with respect to the Wiener filter defined by the above equation (18), the SN ratio estimation unit 18A replaces the speech PSD with the speech model parameter ^{S S} and the noise PSD with the noise model parameter 、 ^N to obtain a gain The coefficient R _{j, k} (ω, l) is calculated by equation (29).

  Hereinafter, points of the noise suppression technique of the present invention will be described. In speech collection under noise, speech PSD and noise PSD can be accurately estimated from spatial information acquired by beamforming. On the other hand, when data similar to the input speech or noise can be prepared in advance, the speech PSD or the noise PSD can be corrected to a more accurate one by constructing a statistical model. Therefore, it is possible to model an observation sound in which noise is mixed in the speech with high accuracy. Here, in the method of designing the Wiener filter by inputting the observation sound to the observation sound model, spatial information is not used and it is difficult to improve the performance, so a specific direction signal acquired by beamforming is input. Thus, by using spatial information and sound source information in combination, a high-performance Wiener filter can be designed, and a high noise suppression amount can be realized without distorting speech.

  The present invention is not limited to the above-described embodiment, and it is needless to say that changes can be made as appropriate without departing from the spirit of the present invention. The various processes described in the above embodiment are not only executed chronologically according to the order described, but may be executed in parallel or individually depending on the processing capability of the apparatus executing the process or the necessity.

[Program, recording medium]
When various processing functions in each device described in the above embodiments are implemented by a computer, the processing content of the function that each device should have is described by a program. By executing this program on a computer, various processing functions in each of the above-described devices are realized on the computer.

  The program describing the processing content can be recorded in a computer readable recording medium. As the computer readable recording medium, any medium such as a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, etc. may be used.

  Further, the distribution of this program is carried out, for example, by selling, transferring, lending, etc. a portable recording medium such as a DVD, a CD-ROM, etc. in which the program is recorded. Furthermore, this program may be stored in a storage device of a server computer, and the program may be distributed by transferring the program from the server computer to another computer via a network.

  For example, a computer that executes such a program first temporarily stores a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. Then, at the time of execution of the process, the computer reads the program stored in its own recording medium and executes the process according to the read program. Further, as another execution form of this program, the computer may read the program directly from the portable recording medium and execute processing according to the program, and further, the program is transferred from the server computer to this computer Each time, processing according to the received program may be executed sequentially. In addition, a configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes processing functions only by executing instructions and acquiring results from the server computer without transferring the program to the computer It may be Note that the program in the present embodiment includes information provided for processing by a computer that conforms to the program (such as data that is not a direct command to the computer but has a property that defines the processing of the computer).

  Further, in this embodiment, although the present apparatus is configured by executing a predetermined program on a computer, at least a part of the processing contents may be realized as hardware.

10 model parameter storage unit 11 microphone array 11-1 to 11-M microphone 12-1 to 12-Q beam former unit 13-1 to 13-Q frequency domain conversion unit 14 specific direction selection unit 15 signal amount estimation unit 16 vector Element extraction unit 17 Specific direction signal modeling unit 18 Gain coefficient calculation unit 19 Multiplication unit 20 Inverse frequency domain conversion unit

Claims (7)

A model parameter storage unit storing parameters of a signal component model in which each output probability of at least one of a clean speech signal and a noise signal and a silent signal is represented by a mixed normal distribution containing a plurality of normal distributions,
It belongs to the angle area signal of the angle area of the desired direction from the plurality of frequency area signals obtained by dividing each of the plurality of angle area signals collected by emphasizing the sound coming from the plurality of angle areas of different directions into a plurality of band components. A specific direction selector for selecting a specific direction frequency domain signal to be
Sound collection area calculated using suppression area signal estimated power calculated using angle area signal of angle area corresponding to predetermined suppression area and angle area signal of angle area corresponding to predetermined sound collection area When a specific direction signal model represented by a mixed normal distribution containing a plurality of normal distributions is generated using at least one of signal estimation powers and parameters of the signal component model, and the specific direction frequency domain signal is observed A specific direction signal modeling unit that calculates a posteriori probability of the specific direction signal model of
A gain coefficient calculation unit which performs weighted addition of gain coefficients for each frequency band calculated using parameters of the signal component model according to the posterior probability of the specific direction signal model;
A multiplication unit for multiplying the signal amount of each corresponding frequency band of the specific direction frequency domain signal with the gain coefficient;
Only including,
Let _j be the state number, k be the distribution number, ω be the frequency bin number, l be the frame number, and a j be the probability that the speech is state j.
The parameters of the above signal component model are composed of an average μ _{j, k} (ω), a variance σ _{j, k} ² (ω), and a mixing weight w _{j, k} ,
The specific direction signal modeling unit
Using the average μ _{j, k} (ω), the variance σ _{j, k} ² (ω), and the mixing weights w _{j, k} of the above signal component model, the average μ ^Y _{j, k} (ω, l) of the specific direction signal model A model parameter generation unit for ^obtaining the variance σ ^Ys _{j, k} ² (ω) and the mixing weight w ^Ys _{j, k} ,
The likelihood of the particular direction frequency domain signal for state j using the mean μ ^Ys _{j, k} (ω, l), the variance σ ^Ys _{j, k} ² (ω), and the mixing weights w ^Ys _{j, k} of the particular direction signal model For degree p (Y _S | j) and distribution k
Find the likelihood p (Y _s | j, k) of the specific direction frequency domain signal and calculate the posterior probability α _j (l) for the state j and the posterior probability β _{j, k} (l) for the distribution k A probability calculation unit,

Including
Noise suppressor. Each output probability of clean speech signal and the silence signal, respectively, and the model parameter storage unit for storing the parameters of the signal component model representing a mixed normal distribution containing a plurality of normal distributions,
It belongs to the angle area signal of the angle area of the desired direction from the plurality of frequency area signals obtained by dividing each of the plurality of angle area signals collected by emphasizing the sound coming from the plurality of angle areas of different directions into a plurality of band components. A specific direction selector for selecting a specific direction frequency domain signal to be
The sum of the signals included in each angular region is the signal amount of that angular region, and the signal amount of each frequency domain signal is multiplied by the inverse matrix of the gain matrix whose elements are parameters obtained from the directivity characteristics for each angular region. A signal amount estimation unit that estimates a signal amount of each angle region ;
Using suppression domain signal estimation power of aggregate signals of angular regions corresponding to Oh et beforehand suppression area defined and the parameters of the signal component model specification represented by Gaussian Mixture containing a plurality of normal distributions A specific direction signal modeling unit that generates a direction signal model and calculates the posterior probability of the specific direction signal model when the specific direction frequency domain signal is observed;
A gain coefficient calculation unit which performs weighted addition of gain coefficients for each frequency band calculated using parameters of the signal component model according to the posterior probability of the specific direction signal model;
A multiplication unit for multiplying the signal amount of each corresponding frequency band of the specific direction frequency domain signal with the gain coefficient;
Noise suppression device comprising a. Each output probability of noise signals and silence signal, respectively, and the model parameter storage unit for storing the parameters of the signal component model representing a mixed normal distribution containing a plurality of normal distributions,
It belongs to the angle area signal of the angle area of the desired direction from the plurality of frequency area signals obtained by dividing each of the plurality of angle area signals collected by emphasizing the sound coming from the plurality of angle areas of different directions into a plurality of band components. A specific direction selector for selecting a specific direction frequency domain signal to be
The sum of the signals included in each angular region is the signal amount of that angular region, and the signal amount of each frequency domain signal is multiplied by the inverse matrix of the gain matrix whose elements are parameters obtained from the directivity characteristics for each angular region. A signal amount estimation unit that estimates a signal amount of each angle region ;
It is represented by Gaussian Mixture containing a plurality of normal distributions with sound collecting area signal estimate power of aggregate signals of Oh et al beforehand corresponding angular region sound collecting area determined and the parameters of the signal component model A specific direction signal modeling unit that generates a specific direction signal model and calculates a posterior probability of the specific direction signal model when the specific direction frequency domain signal is observed;
A gain coefficient calculation unit which performs weighted addition of gain coefficients for each frequency band calculated using parameters of the signal component model according to the posterior probability of the specific direction signal model;
A multiplication unit for multiplying the signal amount of each corresponding frequency band of the specific direction frequency domain signal with the gain coefficient;
Noise suppression device comprising a. The model parameter storage unit stores parameters of a signal component model in which each output probability of at least one of a clean speech signal and a noise signal and a silent signal is represented by a mixed normal distribution including a plurality of normal distributions,
Angle regions of desired directions from a plurality of frequency domain signals obtained by dividing each of a plurality of angle region signals picked up by emphasizing sounds coming from a plurality of angle regions of different directions from a plurality of frequency region signals Selecting a specific direction frequency domain signal belonging to the angular domain signal of
The suppression region signal estimated power calculated using the angle region signal of the angle region corresponding to the predetermined suppression region and the angle region signal of the angle region corresponding to the predetermined sound collection region are calculated by the specific direction signal modeling unit A specific direction signal model represented by a mixed normal distribution containing a plurality of normal distributions is generated using at least one of the sound collection region signal estimated power calculated using the signal component model and the parameter of the signal component model, and the specific direction A specific direction signal modeling step of calculating the posterior probability of the specific direction signal model when a frequency domain signal is observed;
A gain coefficient calculation step in which a gain coefficient calculation unit performs weighted addition of gain coefficients for each frequency band calculated using the parameters of the signal component model according to the posterior probability of the specific direction signal model;
A multiplication step of multiplying the signal coefficients of the corresponding frequency bands of the specific direction frequency domain signal by the multiplication section;
Only including,
Let _j be the state number, k be the distribution number, ω be the frequency bin number, l be the frame number, and a j be the probability that the speech is state j.
The parameters of the above signal component model are composed of an average μ _{j, k} (ω), a variance σ _{j, k} ² (ω), and a mixing weight w _{j, k} ,
The specific direction signal modeling step is
Using the average μ _{j, k} (ω), the variance σ _{j, k} ² (ω), and the mixing weights w _{j, k} of the above signal component model, the average μ ^Y _{j, k} (ω, l) of the specific direction signal model , Variance σ ^Ys _{j, k} ² (ω), mixing weight w ^Ys _{j, k} model parameter generation step,
The likelihood of the particular direction frequency domain signal for state j using the mean μ ^Ys _{j, k} (ω, l), the variance σ ^Ys _{j, k} ² (ω), and the mixing weights w ^Ys _{j, k} of the particular direction signal model Find the likelihood p (Y _s | j, k) of the above specific directional frequency domain signal with respect to degree p (Y _s | j) and distribution k, and calculate posterior probability α _j (l) for state j and distribution k by A probability calculation step of calculating the posterior probability β _{j, k} (l);

Including
Noise suppression method. The model parameter storage unit, each output probability of a clean speech signal and the silence signal, respectively, and the parameters of the signal component model representing a mixed normal distribution containing a plurality of normal distributions is stored,
Angle regions of desired directions from a plurality of frequency domain signals obtained by dividing each of a plurality of angle region signals picked up by emphasizing sounds coming from a plurality of angle regions of different directions from a plurality of frequency region signals Selecting a specific direction frequency domain signal belonging to the angular domain signal of
The signal amount estimation unit takes the sum of the signals included in each angular region as the signal amount of that angular region, and each frequency domain signal with respect to the inverse matrix of the gain matrix whose elements are parameters obtained from the directivity characteristics for each angular region. multiplied by the amount of the signal, and the signal estimation step of estimating a signal amount of each angular region,
Specific direction signal modeling unit contains a plurality of normal distribution using a parameter of the suppression area signal estimation power and the signal component model aggregate signal of the angle region corresponding to the suppression area predetermined Gaussian mixture A specific direction signal modeling step of generating a specific direction signal model represented by a distribution, and calculating a posterior probability of the specific direction signal model when the specific direction frequency domain signal is observed;
A gain coefficient calculation step in which a gain coefficient calculation unit performs weighted addition of gain coefficients for each frequency band calculated using the parameters of the signal component model according to the posterior probability of the specific direction signal model;
A multiplication step of multiplying the signal coefficients of the corresponding frequency bands of the specific direction frequency domain signal by the multiplication section;
Noise suppression method, including. The model parameter storage unit stores parameters of a signal component model in which each output probability of the noise signal and the silence signal is represented by a mixed normal distribution including a plurality of normal distributions,
Angle regions of desired directions from a plurality of frequency domain signals obtained by dividing each of a plurality of angle region signals picked up by emphasizing sounds coming from a plurality of angle regions of different directions from a plurality of frequency region signals Selecting a specific direction frequency domain signal belonging to the angular domain signal of
The signal amount estimation unit takes the sum of the signals included in each angular region as the signal amount of that angular region, and each frequency domain signal with respect to the inverse matrix of the gain matrix whose elements are parameters obtained from the directivity characteristics for each angular region. Signal amount estimation step of estimating the signal amount of each angle region by multiplying the signal amount of
A mixture in which a specific direction signal modeling unit contains a plurality of normal distributions using sound collection area signal estimated power in which signal amounts of angle areas corresponding to predetermined sound collection areas are collected and parameters of the signal component model A specific direction signal modeling step of generating a specific direction signal model represented by a normal distribution, and calculating the posterior probability of the specific direction signal model when the specific direction frequency domain signal is observed;
A gain coefficient calculation step in which a gain coefficient calculation unit performs weighted addition of gain coefficients for each frequency band calculated using the parameters of the signal component model according to the posterior probability of the specific direction signal model;
A multiplication step of multiplying the signal coefficients of the corresponding frequency bands of the specific direction frequency domain signal by the multiplication section;
Noise suppression method. The program for functioning a computer as each part of the noise suppression apparatus in any one of Claim 1 to 3 .

Images