JP5670298B2 - Noise suppression device, method and program - Google Patents

Description

  The present invention relates to a noise suppression technique for extracting a desired signal by suppressing a noise signal included in an input acoustic signal.

  Conventional techniques for suppressing noise signals are known in order to make it easier to hear audio signals from acoustic signals including audio signals to be processed and signals other than audio signals (hereinafter referred to as “noise signals”). In particular, when the automatic speech recognition technology is used in an actual environment, it is necessary to remove a noise signal from an acoustic signal and extract only a desired speech signal in order to correctly perform speech recognition. The use of automatic speech recognition in the actual environment is highly expected in the information-oriented society in the future, and is a problem that should be solved as soon as possible. Non-Patent Documents 1 and 2 are known as conventional techniques related to noise suppression.

  Non-Patent Document 1 discloses the following noise suppression method. The acoustic signal is input, and a probability model of the acoustic signal is generated from the probability models of the speech signal and the noise signal estimated in advance. At this time, the difference between the probability model of each of the speech signal and the noise signal constituting the probability model of the acoustic signal and the statistics of each of the speech signal and the noise signal included in the acoustic signal is expressed by Taylor series approximation. The difference is estimated using an EM algorithm (hereinafter also referred to as “expected value maximization method”), and the probability model of the acoustic signal is optimized. After that, noise is suppressed using the optimized acoustic signal probability model and speech signal probability model parameters.

  Non-Patent Document 2 discloses the following noise suppression method. In order to deal with a noise signal having an acoustic signal as an input and a statistical property following a multimodal distribution, only the noise signal is estimated from the acoustic signal. Using the estimated noise signal, a stochastic model of the noise signal following the multimodal distribution is estimated by the EM algorithm. Thereafter, an optimal probability model of the acoustic signal is generated from the previously estimated probability model of the speech signal and the estimated probability model of the noise signal. Then, noise is suppressed using the optimized parameters of the acoustic signal probability model and the speech signal probability model.

P. J. Moreno, B. Raj, and R. M. Stern, "A vector Taylorseries approach for environment-independent speech recognition", in Proceedings of ICASSP '96, May 1996, vol. II, pp. 733-736 Masayoshi Fujimoto, Tomohiro Nakatani, Shinji Watanabe, “Examination of noise suppression based on robust online estimation of noise model”, IEICE Technical Report, May 2011, SP-2011-2, pp.7-12

  Non-Patent Document 1 is a technique for performing noise suppression on the premise that the characteristics of a noise signal included in an acoustic signal are stationary and the distribution (frequency distribution or probability distribution) is unimodal. However, the characteristics of noise signals in the real environment are non-stationary, and their distribution is often multimodal. For this reason, the technique described in Non-Patent Document 1 cannot cope with non-stationary noise signals, and sufficient noise suppression performance cannot be obtained. In addition, since the relationship between the audio signal and the noise signal included in the acoustic signal is expressed by a nonlinear function, an analytical solution can be obtained when estimating the parameters of the probabilistic models of the audio signal and the noise signal even if Taylor series approximation is used. I can't. For this reason, the technique described in Non-Patent Document 1 cannot obtain the optimal solution of the probability model parameters of each of the speech signal and the noise signal, and cannot obtain sufficient noise suppression performance.

  Non-Patent Document 2 discloses a method of extracting only a noise signal from an acoustic signal and estimating a probability model of a noise signal according to a multimodal distribution using only the extracted noise signal. Therefore, the technique described in Non-Patent Document 2 can deal with non-stationary noise signals. However, the technique described in Non-Patent Document 2 has the following problems. When extracting only a noise signal from an acoustic signal, a stochastic model of an audio signal is required. Generally, a probability model of a speech signal is learned using speech data for learning of a large number of speakers, and this is called a speaker independent model. However, since the statistical properties of speech signals vary greatly depending on the speaker, in order to obtain sufficient noise suppression performance, the speech data for learning of a specific speaker is not used as the probability model of the speech signal. It is necessary to use a speaker-dependent model learned by using the speaker-dependent model or a speaker-dependent model in which the speaker-independent model is adapted to the statistics of a specific speaker. However, the technique described in Non-Patent Document 2 does not consider such a speaker dependence model, and does not provide sufficient noise suppression performance.

  Also, in order to estimate the stochastic model of the noise signal and adapt the speaker independent model to the statistic of the speech signal of the specific speaker, learning data of only the noise signal and the speech signal is required respectively. The signal that can be observed when performing noise suppression is only a signal in which a noise signal and a voice signal are mixed, and it is impossible to observe only the noise signal and the voice signal alone. For this reason, in Non-Patent Document 1, a time interval in which only a noise signal or a voice signal exists is estimated from an acoustic signal, and learning data of only the noise signal or the voice signal is obtained. However, in such a method, a noise signal in a time interval in which a sound signal exists or a sound signal in a time interval in which a noise signal exists cannot be used as learning data. For this reason, changes and characteristics of the noise signal and the voice signal generated in the section cannot be reflected in the noise signal probability model and the speaker dependence model. Therefore, in Non-Patent Document 1, it is difficult to accurately estimate the multimodal distribution of noise signals and adapt the speaker independent model to the statistic of the speech signal of a specific speaker. On the other hand, Non-Patent Document 2 does not consider the speaker dependence model as described above.

  The present invention estimates a noise signal and a speech signal included in an acoustic signal, uses the noise signal as learning data regardless of the presence or absence of the speech signal, reflects it in the probability model of the noise signal, and the presence or absence of the noise signal Regardless of the system, the speech signal is used as learning data, and the speaker independent model is reflected in the speaker adaptation parameter, which is a parameter for adapting to the statistic of the speech signal of a specific speaker. It is an object of the present invention to provide a noise suppression technique capable of effectively suppressing a noise signal from an acoustic signal by using a noise signal probability model and a speaker dependence model obtained by speaker adaptation processing.

  In order to solve the above problems, according to the first aspect of the present invention, a noise signal is suppressed from an acoustic signal including a noise signal and a voice signal. Extract the acoustic features of the acoustic signal. This is a stochastic model of speaker-independent speech signals that are defined as non-stationary noise-based signals that follow a multimodal distribution, and that contain many speakers' speech signals as learning data. Using a speaker adaptation parameter for adapting a speaker independent speech model to the speaker of the speech signal included in the acoustic signal, a noise model that is a stochastic model of the noise signal, and a speaker independent speech model A first probability model that is a signal probability model is generated. A noise signal is estimated based on the first probability model and the acoustic features of the acoustic signal, and the noise model is unsupervised and learned using the estimated noise signal as learning data. Using the acoustic features of the acoustic signal, the speaker independent speech model, and the noise model, the speech signal included in the acoustic signal is estimated, and the speaker adaptation parameter is unsupervised by using the estimated speech signal as learning data. The noise signal contained in the acoustic signal is suppressed using the acoustic feature of the acoustic signal, the speaker independent speech model, the noise model, and the speaker adaptation parameter.

  The noise suppression technology according to the present invention estimates a noise signal and a speech signal included in an acoustic signal, estimates a probabilistic model of a multimodal noise signal using the estimated noise signal, and uses the estimated speech signal By estimating the speaker adaptation parameters, it is possible to effectively suppress the noise signal from the acoustic signal and extract the target speech signal even in an environment where various noises exist.

The functional block diagram of the noise suppression apparatus of 1st embodiment. The figure which shows the processing flow of the noise suppression apparatus of 1st embodiment. The figure which shows the processing flow of an acoustic feature-value extraction part. The functional block diagram of a parameter estimation part. The figure which shows the processing flow of a parameter estimation part. The functional block diagram of a noise model estimation part. The figure which shows the processing flow of a noise model estimation part. The functional block diagram of a speaker adaptive parameter estimation part. The figure which shows the processing flow of a speaker adaptive parameter estimation part. The functional block diagram of a noise suppression part. The figure which shows the processing flow of a noise suppression filter estimation means. The figure which shows the processing flow of a noise suppression filter application means. The figure which shows the simulation result of the noise suppression apparatus of 1st embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings used for the following description, constituent parts having the same function and steps for performing the same processing are denoted by the same reference numerals, and redundant description is omitted. In the following description, the symbols “^”, “ ⁻ ”, etc. used in the text should be described immediately above the character that immediately follows, but are described immediately before the character due to restrictions on the text notation. To do. In the formula, these symbols are written in their original positions. Further, the processing performed for each element of a vector or matrix is assumed to be applied to all elements of the vector or matrix unless otherwise specified.

<First embodiment>
In the present embodiment, a Gaussian Mixture Model (hereinafter also referred to as “GMM”) is adopted as a noise signal probability model and speaker-dependent model based on a multimodal distribution.

As shown in FIG. 1, the noise suppression apparatus 100 includes an acoustic feature extraction unit 104, a GMM storage unit 107 that stores a silence GMM and a clean speech GMM that constitute a speaker independent speech model, a parameter estimation unit 105, Noise suppression unit 106. The noise suppression apparatus 100 records or inputs an acoustic signal o _τ obtained by mixing the audio signal s _τ and the noise signal n _τ, and a noise suppression signal ^ s obtained by suppressing the noise signal n _τ from the acoustic signal o _{τ. τ} is output. However, (tau) represents the sample point of a discrete signal. Hereinafter, an outline of the present embodiment will be described.

Acoustic feature extraction unit 104 as shown in FIG. 2 extracts the complex spectrum Spc _t and logarithmic Mel spectrum O _t is a feature amount for performing noise suppression of acoustic signals o τ _(s104).

The parameter estimation unit 105 includes a logarithmic mel spectrum O _t and a speaker independent (Speaker Independent, hereinafter referred to as “SI”) speech signal probability model (hereinafter referred to as “SI speech model”) held in the main memory by the GMM storage unit 107. SIGMM parameter set λ _SI is used, noise signal probability model (hereinafter referred to as “noise model”) noise GMM parameter set λ _N , and SIGMM is the audio signal s included in the acoustic signal o _τ. The speaker adaptation parameter b for adapting to the speaker of _τ is estimated (s105).

The noise suppression unit 106 designs a noise suppression filter using the complex spectrum Spc _t , the log mel spectrum O _t , the SIGMM parameter set λ _SI , the noise GMM parameter set λ _N, and the speaker adaptation parameter b, The noise signal n _τ is suppressed from o _τ to obtain a noise suppression signal ^ s _τ (s106). Details of each part will be described below.

<Acoustic Feature Extraction Unit 104>
Acoustic feature extraction unit 104 receives the acoustic signal o _tau, extracted acoustic features of the audio signal o _tau (s104), and outputs to the parameter estimation unit 105 and the noise suppression unit 106. The extracted acoustic features are used when a noise signal is suppressed from an acoustic signal, and are, for example, a complex spectrum and a log mel spectrum. For example, the acoustic feature extraction unit 104 performs processing according to the flow shown in FIG.

First, while moving the start point at a certain frequency (e.g. 16,000Hz) in the sampled sound signal o _tau the time axis direction in a predetermined time width (shift width), the frame acoustic signal for a predetermined time length (frame width) Is cut out (s201). For example acoustic signals of the frame width Frame = 320 samples points _{(16,000Hz × 20ms) o t =} {o t, 0, o t, 1, ..., o t, n, ..., o t, Frame-1} Is cut out while shifting the start point by shift width Shift = 160 sample points (16,000 Hz × 10 ms). Here, t represents the frame number, and n represents the nth sample point in the frame. Note that when an acoustic signal of a plurality of channels is input, a frame may be cut out for each channel. Further, when cutting out the frame, it may be excised for example by multiplying the window function w _n, such as the following Hamming window.

Next, the acoustic feature extraction unit 104 applies an M-point fast Fourier transform process to the acoustic signal ot _{, n} to thereby calculate a complex spectrum Spc _t = {Spc _{t, 0} ,..., Spc _{t, m} ,. Spc _{t, M-1} } is obtained (s202). However, M must be set to a power of 2 and a value equal to or larger than the frame width Frame, for example, 512. M is the frequency bin number.

Next, the acoustic feature extraction unit 104 performs mel filter bank analysis on the absolute value of Spct _{, m} (s203), and applies logarithmic processing to the output of the filter bank (s204). By such processing, a vector having an R-dimensional (for example, R = 24) log mel spectrum as an element (hereinafter, this vector is simply referred to as “log mel spectrum”) O _t = {O _{t, 0} ,..., O _{t , R 1} ,..., O _{t, R−1} } is calculated. However, r shows the element number of a vector. That is, the output of the acoustic feature extraction unit 104 is the complex spectrum Spt _t and the log mel spectrum O _t . Complex spectrum Spc _t becomes the input of the noise suppressor 106, logarithmic Mel spectrum _{O t} is a parameter estimation unit 105, the input of the noise suppressor 106.

<GMM storage unit 107>
A storage unit (not shown) stores in advance SIGMMs that are learned using voice signals of many speakers without learning noise as learning data. The SIGMM consists of a silent GMM and a clean voice GMM. The silent GMM is a GMM learned based on an acoustic signal acquired from a silent part of a voice signal that does not include a noise signal, and the clean voice GMM is based on an acoustic signal consisting only of a voice excluding the silent part in a noise-free environment. It is a learned GMM.
SIGMM is given by:

Here, j is an index for identifying the silent GMM and the clean voice GMM, j = 0 indicates the silent GMM, and j = 1 indicates the clean voice GMM. Further, k is a normal distribution number included in the silent GMM or the clean speech GMM, and K is the total normal distribution number (for example, K = 128). Further, _{S t} is free of noise, the logarithm Mel spectrum of the speech _{signal, b SI, j (S t} ) is the likelihood of SIGMM. w _{SI, j, k} and μ _{SI, j, k} and ΣSI _{, j, k} are the SIGMM mixture weight, average vector and diagonal dispersion matrix, respectively _, and are preliminarily used by learning speech data for many speakers. To estimate. A SIGMM parameter set (hereinafter also referred to as “SI model parameter set”) is defined as λ _SI = {w _{SI, j, k} , μ _{SI, j, k} , ΣSI _{, j, k} }. The subscript SI indicates a likelihood or parameter related to SIGMM. The function N (•) is a probability density function of a multidimensional normal distribution given by the following equation.

In the above formula, “ ^T ” represents transposition.
Hereinafter, a speaker-dependent GMM (Speaker Dependent GMM, hereinafter referred to as “SDGMM”), which is a speaker-dependent model used in the present embodiment, and a noise GMM, which is a noise model, will be outlined. The SDGMM is a GMM composed of parameters estimated from voice data for learning of a specific speaker. As described above, the speaker dependence model includes a model learned using speech data for learning of a specific speaker, or a model in which a speaker independent model is adapted to the statistics of the specific speaker. Obtaining an SDGMM from only speech data for learning of a specific speaker is impractical in practice, and in this embodiment, an SDGMM is obtained by performing speaker adaptation processing on the SIGMM. That is, the SIGMM average vector μSI _{, j, k} is converted into the SDGMM average vector μSD _{, j, k} by the speaker adaptation processing of the following equation.

In the above equation, b is a speaker adaptation parameter composed of an R-dimensional vector, and is an independent parameter for j and k. The subscript SD indicates the likelihood and parameters related to SDGMM.
The noise GMM is given by the following equation.

In the above equation, l is a normal distribution number included in the noise GMM, and L is the total normal distribution number (for example, L = 4). N _t is a log mel spectrum of the noise signal, and b _N (N _t ) is the likelihood of the noise GMM. w _{N, l} and μ _{N, l} and ΣN _{, l} are the noise GMM mixing weight, average vector and diagonal dispersion matrix, respectively. Hereinafter, a noise GMM parameter set (hereinafter also referred to as a “noise model parameter set”) is defined as λ _N = {w _{N, l} , μ _{N, l} , Σ _{N, l} }. Note that the subscript N indicates the likelihood or parameter related to the noise GMM.

  In Non-Patent Document 1, noise suppression is performed on the assumption that the characteristics of a noise signal are stationary and the distribution is unimodal. On the other hand, in the present embodiment, the noise signal is defined as a signal based on non-stationary noise following a multimodal distribution, and the noise model is expressed by GMM instead of a single normal distribution. The parameter estimation unit 105 described later performs unsupervised learning of the noise model.

  In Non-Patent Document 2, the speaker dependence model is not used. On the other hand, in this embodiment, speaker adaptation processing is performed on the SIGMM to obtain an SDGMM, and a speaker dependence model is used in noise suppression. Note that an adaptive parameter is estimated by a parameter estimation unit 105 described later.

<Parameter estimation unit 105>
The parameter estimation unit 105 receives the log mel spectrum O _t and the SI model parameter set λ _SI, and uses these values to optimally estimate the noise model parameter set λ _N and the speaker adaptation parameter b (s105), and noise. The result is output to the suppression unit 106.

In the parameter estimation unit 105, the noise model parameter set λ _N and the speaker adaptation parameter b are estimated by three types of nested EM algorithms. Hereinafter, they are referred to as a first EM algorithm, a second EM algorithm, and a third EM algorithm. The EM algorithm is a method used for parameter estimation of a certain probability model, and expectation-step (E-step) for calculating the expected value of the cost function (log likelihood function) of the probability model and maximizing the cost function. The parameter is optimally estimated by repeating Maximization-step (M-step) to satisfy the convergence condition.

  The parameter estimation unit 105 includes a first initial value estimation unit 302, a first probability model generation unit 303, a first expected value calculation unit 304, a noise model estimation unit 305, a speaker adaptive parameter estimation unit 306, and a first convergence determination unit 307. (See FIG. 4). The processing flow of the parameter estimation unit 105 will be described with reference to FIG.

(First initial value estimation means 302)
The first initial value estimating means 302 initializes the repetition index i of the first EM algorithm (s301). For example, i = 1. Further, the first initial value estimating means 302 receives the logarithmic mel spectrum O _t of the acoustic signal o _t and uses this value to set the initial value λ _N ^{(i = 0)} of the noise model parameter set λ _N in the first EM algorithm. {W _{N, l} ^{(i = 0)} , μ _{N, l} ^{(i = 0)} , Σ _{N, l} ^{(i = 0)} } are estimated by the following equations (6) to (8), and the speaker adaptation parameters The initial value of b is set by the following equation (9) (s302) and output to the first probability model generation means 303.

In the above formula, ^- 0 vector of all elements 0, U is the number of frames required for the initial value estimation (e.g. U = 10). “diag” indicates a diagonal matrix having elements in parentheses, and the superscript (i) indicates a parameter in the i-th iteration estimation in the first EM algorithm.

(First probability model generation means 303)
The first probability model generation unit 303 obtains the noise model parameter set λ ⁽ⁱ⁻¹⁾ _N and the speaker adaptation parameter b ^{(i−1) in the i−} 1th iterative estimation from the first convergence determination unit 307 described later. receive. However, if the i−1th noise model parameter set λ _N ⁽ⁱ⁻¹⁾ and the speaker adaptation parameter b ⁽ⁱ⁻¹⁾ do not exist, that is, in the 0th time, the initial value λ _N ^{(i = 0)} and b ^{(i = 0)} are received from the first initial value estimating means 302. Further first probability model generation unit 303 receives the SI model parameter set lambda _SI, using these values, the probability model of the logarithmic Mel spectrum O _t of the audio signal o _t (hereinafter referred to as "first probability model") Is composed of the following GMM.

In the above equation, b _{O1, j} (O _t ) is the likelihood of the probability model of the log mel spectrum O _t , the function N (•) is given by equation (3), and w _{O1, j, k, l} ^{(i )} , Μ _{O1, j, k, l} ⁽ⁱ⁾ and _{ΣO1, j, k, l} ⁽ⁱ⁾ are the noise model parameter set λ _N ^(i-1) and the speaker in the ⁱ⁻ 1th iteration estimation. The mixture weight, average vector, and diagonal dispersion matrix of the probability model of the log mel spectrum O _t generated from the adaptive parameter b ⁽ⁱ⁻¹⁾ and the SI model parameter set λ _SI , and the following equation (11) It is given by (13).

In the above equation, the functions log (•) and exp (•) perform an operation for each element r of the vector. Also, ^- 1 vector of all elements 1, I is the identity ^{_{matrix, H j, k, l (}} i) is the Jacobian matrix of the function h (·).

A parameter set of the first probability model (hereinafter referred to as “first probability model parameter set”) is expressed as λ _O1 ⁽ⁱ⁾ = {w _{O1, j, k, l} ⁽ⁱ⁾ , μ _{O1, j, k, l} ⁽ⁱ⁾ , _{ΣO1, j, k, l} ⁽ⁱ⁾ }.

The first probability model generation unit 303 generates λ _O1 ⁽ⁱ⁾ based on the equations (11) to (13) (s303), and outputs it to the first expected value calculation unit 304 and the noise model estimation unit 305.
(First expected value calculation means 304)
First expectation value calculation unit 304 receives the first probabilistic model parameter set lambda _{O1 ⁽ⁱ⁾} and a logarithmic Mel spectrum O _t of the audio signal o _t, the probability model of the logarithmic Mel spectral O _t in the i-th repetition estimate The expected value of the cost function Q ₁ (•) is calculated by the following equation (E-step of the first EM algorithm) (s304).

In the above equation, O _{0: T−1} = {O ₀ ,..., O _t ,..., O _T−1 }, T is the total number of frames of the log mel spectrum O _t , and P _{t, j} ^{(i )} Is a posterior probability for GMM type j in frame t given by equation (16). In particular, P _{t, j = 0} ⁽ⁱ⁾ is defined as a speech non-existence probability, and P _{t, j = 1} ⁽ⁱ⁾ is defined as a speech presence probability. P _{t, j, k, l} ⁽ⁱ⁾ is the posterior probability for normal distribution numbers k and l in frame t given by equation (17). λ _O1 ⁽ⁱ⁾ = {w _{O1, j, k, l} ⁽ⁱ⁾ , _{μO1, j, k, l} ⁽ⁱ⁾ , _{ΣO1, j, k, l} ⁽ⁱ⁾ } 13).

The first expected value calculation unit 304 uses the obtained first expected value Q ₁ as the first convergence determination unit 307 and the posterior probabilities P _{t, j} ⁽ⁱ⁾ and P _{t, j, k, l} ⁽ⁱ⁾ as noise models. It outputs to the estimation means 305.

  The M-step of the first EM algorithm includes the following processing in the noise model estimation unit 305 and the speaker adaptive parameter estimation unit 306. The second EM algorithm is implemented in the noise model parameter estimation means 402 in the noise model estimation means 305, and the third algorithm is implemented in the speaker adaptation parameter update means 503 in the speaker adaptation parameter estimation means 306.

(Noise model estimation means 305)
Noise model estimating means 305, logarithmic Mel spectrum _{O t} and the posterior probability _P ^t of the first probabilistic model parameter set λ _O1 ⁽ⁱ⁾ an acoustic signal _{o t, j (i),} P t, j, k, l (i ^{) And} using these values, the log mel spectrum N _t of the noise signal is estimated, and the noise model ^ N _t is used as learning data to train the noise model unsupervised (s305), and the noise model parameter set λ _N ⁽ⁱ⁾ is output to speaker adaptive parameter estimation means 306 and first convergence determination means 307. The noise model estimation unit 305 includes a noise signal estimation unit 401 and a noise model parameter estimation unit 402 (see FIG. 6).

[Noise signal estimation means 401]
The noise signal estimation means 401 estimates the log mel spectrum N _t of the noise signal using the log mel spectrum O _t (s401 in FIG. 7). For example, the noise signal estimation unit 401 uses the posterior probabilities P _{t, j} ⁽ⁱ⁾ , P _{t, j, k, l} ⁽ⁱ⁾ and the average vector μ included in the first probability model parameter set λ _O1 ⁽ⁱ⁾ ( _{O1, j = 1, k, l} ⁽ⁱ⁾ ), the log mel spectrum O _t, and the noise vector parameter set λ _N ⁽ⁱ⁻¹⁾ (in the ⁱ⁻ 1th iteration estimation, the average vector μ _N ^{( i-1)} ) and using these values, the log mel spectrum N _t ⁽ⁱ⁾ of the noise signal used to update the noise model parameter set λ _N ^{(i-1 )} is estimated, and the estimated value ^ N _t ⁽ⁱ⁾ is output to the noise model parameter estimation means 402. The log mel spectrum N _t ⁽ⁱ⁾ of noise is estimated by the following equation.

[Noise model parameter estimation means 402]
The noise model parameter estimator 402 receives the estimated log mel spectrum value {circumflex over ( _N ⁾ } _t ⁽ⁱ⁾ of the noise signal, and estimates the noise model parameter set λ _N ⁽ⁱ⁾ as learning data (from s402 to s402 in FIG. 7 ⁾ . s407), and outputs the result to the first convergence determination unit 307 and the speaker adaptation parameter estimation unit 306. A specific estimation method (s402 to s407) of the noise model parameter set λ _N ⁽ⁱ ) will be described later.

(Speaker adaptive parameter estimation means 306)
Speaker adaptation parameter estimation section 306 receives the log-Mel spectrum O _t and SI model parameter set lambda _SI and noise model parameter set λ _{N ^(i),} using these values, the speech contained in the audio signal o _t estimating the signal _{s t,} as learning data the estimated speech signal ^ _{s t,} speaker adaptation parameter ^{b (i)} and the estimated unsupervised (s306 in Fig. 5) to the first convergence determining means 307.

The speaker adaptation parameter estimation unit 306 includes second probability model generation unit 501, speech signal estimation unit 502, and speaker adaptation parameter update unit 503 (see FIG. 8).
[Second probability model generation means 501]
The second probability model generation means 501 receives the noise model parameter set λ _N ⁽ⁱ⁾ , the speaker adaptation parameter b ^(i−1), and the SI model parameter set λ _SI, and uses these values to calculate the logarithmic mel. A probability model (hereinafter referred to as “second probability model”) of the spectrum O _t is configured by the following GMM.

In the above equation, b _{O2, j} (O _t ) is the likelihood of the probability model of the log mel spectrum O _t , the function N (•) is given by equation (3), and w _{O2, j, k, l} ^{( i)} , μ _{O2, j, k, l} ⁽ⁱ⁾ , ΣO2 _{, j, k, l} ⁽ⁱ⁾ are the mixture weight, average vector, and diagonal dispersion matrix of the second probability model, 20) to (22).

A parameter set of the second probability model (hereinafter referred to as “second probability model parameter set”) is expressed as λ _O2 ⁽ⁱ⁾ = {w _{O2, j, k, l} ⁽ⁱ⁾ , _{μO2, j, k, l} ⁽ⁱ⁾ , ΣO2 _{, j, k, l} ⁽ⁱ⁾ }.
The second probability model generation means 501 generates a second probability model parameter set λ _O2 ⁽ⁱ⁾ based on the equations (20) to (22) (s501 in FIG. 9), and λ _O2 ⁽ⁱ⁾ and the log mel spectrum O _t and the SI model parameter set λ _SI are output to the speech signal estimating means 502.

[Speech signal estimation means 502]
The speech signal estimation means 502 uses the log mel spectrum O _t , the SI model parameter set λ _SI (the average vector μ _{SI, J, k} ) included in the second probability model parameter set λ _O2 ^(i), and the i−1th time. receiving a speaker adaptation in the repeating estimated parameter b ^(i-1), using these values, is used to update the speaker adaptation parameter b ^(i), the clean speech of logarithm contained in the audio signal o _t The mel spectrum S _t is estimated by the following equation (s502), and the estimated value ｔS _t ⁽ⁱ⁾ is output to the speaker adaptive parameter updating unit 503.

Incidentally, the posterior probability ^{_{P t, j (i),}} P t, j, k, l (i) , the second probability model parameter set lambda _O2 instead of the first probabilistic model parameter set λ _O1 ^{(i) (i)} Is given by equations (16) and (17).

[Speaker adaptive parameter update means 503]
The speaker adaptation parameter update means 503 receives the log mel spectrum ^ S _t ⁽ⁱ⁾ of the estimated clean speech and the SI model parameter set λ _SI, and uses these values to determine the speaker adaptation parameter b ^(i−1). Is updated (s503 to s508 in FIG. 9) and output to the first convergence determination means 307. A specific updating method (s503 to s508) of the speaker adaptation parameter b ⁽ⁱ ) will be described later.

(First convergence determination means 307)
First convergence determining means 307 receives the first expected value _{Q 1,} to determine the convergence condition is satisfied whether using this value (s307 in Fig. 5), if it meets λ _N = λ _N ⁽ⁱ⁾ , B = b ⁽ⁱ⁾ , λ _N and b are output to the noise suppression unit 106, and the processing of the parameter estimation unit 105 ends. If not satisfied, λ _N ⁽ⁱ⁾ and b ⁽ⁱ⁾ are output to the first probability model generation means 303, and as i ← i + 1 (s308), a control signal is output to each unit so that iterative processing is performed. The process of s306 is repeated. For example, the convergence condition includes the latest first expected value Q ₁ (O _{0: T−1} , b ⁽ⁱ⁾ , λ _N ⁽ⁱ⁾ ) and the previous first expected value Q ₁ (O _{0: T− 1} , b ⁽ⁱ⁻¹⁾ , λ _N ⁽ⁱ⁻¹⁾ ) is less than or equal to a predetermined value η _1, or the number of repetitions i is greater than or equal to a predetermined value I _1. . For example, it can be expressed by the following formula.

For example, η ₁ = 0.0001 and I ₁ = 100.
<Details of Noise Model Parameter Estimation Unit 402>
The noise model parameter estimation unit 402 includes a second initial value estimation unit 403, a second expected value calculation unit 404, a parameter update unit 405, and a second convergence determination unit 406 (see FIG. 6). The processing contents of the noise model parameter estimation unit 402 will be described with reference to FIG.

(Second initial value estimating means 403)
The second initial value estimation means 403 first initializes an index i ′ indicating the number of repetitions of the second EM algorithm (s402). For example, i ′ = 1. Next, the second initial value estimation means 403 receives the logarithmic mel spectrum ^ N _t ⁽ⁱ⁾ of the estimated noise signal, and uses this value for the noise model parameter set λ _N ^{(i ′)} in the second EM algorithm. Initial values λ _N ^{(i ′ = 0)} = {w _{N, l} ^{(i ′ = 0)} , μ _{N, l} ^{(i ′ = 0)} , Σ _{N, l} ^{(i ′ = 0)} } To (30) (s403) and output to the second expected value calculation means 404.

In the above equation, the superscript (i ′) indicates a parameter in the i′-th iteration estimation. GaussRand (a, b) is a generator of normal random numbers having an average a and a variance b.

(Second expected value calculation means 404)
The second expected value calculation unit 404 receives the logarithmic mel spectrum ^ N _t ⁽ⁱ⁾ of the estimated noise signal from the noise signal estimation unit 401. Also, the noise model parameter set λ _N ^(i′−1) in the second EM algorithm is received from the second convergence determination means 406. However, if the i′-1th noise model parameter set λ _N ^(i′−1) does not exist, that is, in the 0th time, the above-mentioned initial value λ _N ^{(i ′ = 0)} is used as the second initial value estimation. Receive from means 403. Using these values, the second expected value calculation means 404 calculates the expected value of the cost function Q ₂ (•) of the noise GMM in the i′-th iterative estimation according to the equation (31) (E− of the second EM algorithm). step) (s404) and output to the second convergence determination means 406.

In the above equation, ^ N _{0: T-1} ⁽ⁱ⁾ = {^ N ₀ ⁽ⁱ⁾ , ..., ^ N _t ⁽ⁱ⁾ , ..., ^ N _T-1 ⁽ⁱ⁾ },
The function N (•) is given by equation (3), and P _{t, l} ^{(i ′)} is the posterior probability for the normal distribution number l in the frame t given by the following equation.

The second expected value calculation unit 404 outputs the obtained P _{t, l} ^{(i ′)} to the parameter update unit 405.
(Parameter update means 405)
The parameter updating means 405 receives the posterior probability P _{t, l} ^{(i ′)} and the log mel spectrum ^ N _t ^(i), and updates the noise model parameter set λ _N ^(i′−1) by the following equation (first The second EM algorithm M-step) (s405) and the updated noise model parameter set λ _N ^{(i ′)} are output to the second convergence determination means 406.

(Second convergence determination means 406)
Second convergence determining means 406 receives the second expected value _{Q 2,} to determine the convergence condition is satisfied whether using this value (s406), if it meets ^{_{λ N (i) = λ N}} (i ' ⁾ And λ _N ⁽ⁱ⁾ are output to the first convergence determination unit 307 and the speaker adaptation parameter update unit 503, and the process of the noise model parameter estimation unit 402 is terminated. If not satisfied, λ _N ^{(i ′)} is output to the second expected value calculation means 404, and as i ′ ← i ′ + 1 (s407), a control signal is output to each part so as to perform repeated processing, and s404, s405 Repeat the process. For example, the convergence condition includes the latest second expected value Q ₂ (^ N _{0: T−1} , λ _N ^{(i ′)} ) and the previous second expected value Q ₂ (^ N _{0: T−1} , The difference from λ _N ^(i′−1) ) may be a predetermined value η ₂ or less, or the number of repetitions i ′ may be a predetermined value I ₂ or more. For example, it can be expressed by the following formula.

For example, η ₂ = 0.0001 and I ₂ = 100.
<Details of Speaker Adaptive Parameter Update Unit 503>
The speaker adaptation parameter update unit 503 includes an initial value setting unit 504, a third expected value calculation unit 505, a speaker adaptation parameter calculation unit 506, and a third convergence determination unit 507 (see FIG. 8).

The processing contents of the speaker adaptation parameter update unit 503 will be described with reference to FIG.
(Initial value setting means 504)
The initial value setting means 504 first initializes an index i ″ indicating the number of repetitions of the third EM algorithm (s503). For example, i ″ = 1. Next, the initial value setting means 504 sets the initial value b ^{(i ″ = 0)} of the speaker adaptation parameter b ^{(i ″)} in the third EM algorithm by the following equation (s504), and the third expected value calculation means 505. Output to.

In the above formula, the superscript (i ″) indicates that it is a parameter in i ″ th iteration estimation in the third EM algorithm.
(Third expected value calculation means 505)
The third expected value calculation means 505 receives the SI model parameter set λ _SI and the estimated log mel spectrum Ｓ S _t ^{(i) of} the clean speech. Also, the i "-1th speaker adaptation parameter b ^{(i" -1)} is received from the third convergence determination means 507. However, if the i "-1th speaker adaptation parameter b ^{(i" -1)} does not exist, that is, if it is the 0th time, the initial value b ^{(i "= 0)} is received from the initial value setting means 504. Using these values, the third expected value calculation means 505 calculates the expected value of the cost function Q ₃ (•) of the SDGMM in the i ”th iterative estimation by the following equation (E-step of the third EM algorithm): (S505), it outputs to the third convergence determination means 507.

In the above _{^{equation, ^ S 0: T-1}} (i) = {^ S 0 (i), ..., ^ S t (i), ..., ^ S T-1 (i)} _{is, P t, j} ^{(I ″)} is the posterior probability for GMM type j in frame t given by the following equation (39), and P _{t, j, k} ^{(i ″)} is the normality in frame t given by the following equation (40). This is the posterior probability for distribution number k.

The third expected value calculation unit 505 outputs the obtained posterior probabilities P _{t, j} ^{(i ″)} and P _{t, j, k} ^{(i ″)} to the speaker adaptive parameter calculation unit 506.

(Speaker adaptive parameter calculation means 506)
The speaker adaptive parameter calculation means 506 includes the posterior probabilities P _{t, j} ^{(i ″)} , P _{t, j, k} ^{(i ″)} , the SI model parameter set λ _SI, and the log mel spectrum of clean speech ^ S _t ^(i). The speaker adaptation parameter b ^{(i ″)} is obtained using the following equation (41), and is updated as a new speaker adaptation parameter b ^{(i ″)} (M-step of the third EM algorithm ^). ) (S506), it outputs to the third convergence determination means 507.

(Third convergence determination means 507)
Third convergence determining unit 507 receives the third expected value _{Q 3,} to determine the convergence condition is satisfied whether using this value (s 507), if it meets the ^{b (i) = b (i} ") b ⁽ⁱ⁾ is output to the first convergence determination unit 307, and the processing of the speaker adaptive parameter update unit 503 is terminated. If not satisfied, b ^{(i ")} is output to the third expected value calculation unit 505, and i" As i ″ +1 (s508), a control signal is output to each unit so as to repeat the process, and the processes of s505 and s506 are repeated. For example, the convergence condition includes the latest third expected value Q ₃ (^ S _{0: T−1} ⁽ⁱ⁾ , b ^{(i ″)} ) and the previous third expected value Q ₃ (^ S _{0: T− 1} ⁽ⁱ⁾ , b ^{(i ″ −1)} ) is a predetermined value η ₃ or less, or the number of repetitions i ″ is a predetermined value I ₃ or more. It can be expressed by the following formula.

For example, η ₃ = 0.0001 and I ₃ = 100. Therefore, the speaker adaptation parameter b ⁽ⁱ⁾ = b ^{(i ″)} finally calculated by the speaker adaptation parameter calculation means 506 in the M-step of the third EM algorithm is the i-th iteration of the first EM algorithm. The likelihood of SDGMM is maximized.

<Noise Suppression Unit 106>
Noise suppressor 106 receives the acoustic signal o logarithmic Mel spectrum is an acoustic characteristic of _t O _t and SI model parameter set lambda _SI and noise model parameter set lambda _N and speaker adaptation parameters b, using these values sound signal _{o t} and suppressing noise signals _{n t} contained in (s106), and outputs a noise suppressed signal ^ s _tau as an output value of the noise suppression apparatus 100. For example, as shown in FIG. 10, the noise suppression unit 106 includes a noise suppression filter estimation unit 601 and a noise suppression filter application unit 602. The noise suppression filter estimation means 601 receives the log mel spectrum O _t of the acoustic signal, the SI model parameter set λ _SI , the noise model parameter set λ _N and the speaker adaptation parameter b, and estimates the noise suppression filter W ^Lin _{t, m} . . Noise suppression filter application unit 602 receives the complex spectrum Spc _t and the noise suppression filter ^W _{Lin t,} and _m, to obtain a noise suppression signal ^ s _tau to suppress noise. Details of each means will be described below.

(Noise suppression filter estimation means 601)
The noise suppression filter estimation means 601 performs processing according to the flow shown in FIG. First, the noise suppression filter estimation means 601 receives the SI model parameter set λ _SI , the noise model parameter set λ _N, and the speaker adaptation parameter b, and uses these values to create a probability model of the log mel spectrum O _t of the acoustic signal. Λ _O3 = {w _{O3, j, k, l} , _{μO3, j, k, l} , _{ΣO3, j, k,} λ _O3 = {w _{O3, j, k, l} , μ _{O3, j, k, l l} } is generated as follows (s601).

In the above equation, the functions h (•) and g (•) are given by equations (12) and (13).

Next, the noise suppression filter estimation means 601 uses the obtained third probability model parameter set λ _O3 and the log mel spectrum O _t and uses the posterior probabilities P _{t, j} , P _{t according to} equations (48) and (49). _{, J, k, l} are calculated (s602).

Next, the noise suppression filter estimation unit 601 uses the SI model parameter set λ _SI , the noise model parameter set λ _N, and the posterior probabilities P _{t, j} , P _{t, j, k, l} on the mel frequency axis. The noise suppression filter W ^Mel _{t, r} is estimated as in the following equation (s603).

The above expression is a notation for each element r of the vector.

Next, the noise suppression filter estimation means 601 converts the noise suppression filter W ^Mel _{t, r} on the mel frequency axis into the noise suppression filter W ^Lin _{t, m} on the linear frequency axis (s604), and the noise suppression filter. The data is output to the application unit 602. Note that the conversion is performed by estimating the value of the noise suppression filter on the linear frequency axis by applying cubic spline interpolation to the mel frequency axis.
(Noise suppression filter applying means 602)
The noise suppression filter application unit 602 performs processing according to the flow shown in FIG. The noise suppression filter application unit 602 receives the noise suppression filter W ^Lin _{t, m} and the complex spectrum Spc _t, and multiplies the complex spectrum Spct _t by the noise suppression filter W ^Lin _{t, m} as follows: A noise-suppressed complex spectrum {circumflex over (S)} _{t, m} is obtained (s701).

The above expression is a notation for each element m of the vector.

Next, the noise suppression filter application unit 602 obtains the complex spectrum _{^ S t,} by applying the inverse fast Fourier transform on _m, the noise suppression signal _{^ s t} in frame _t, the _n (s702).

Next, the noise suppression filter application unit 602 obtains a noise suppression signal ^ s _tau continuous noise suppression signal ^ s t of each _frame, the _n linked with releasing the window function w _n by the following equation (S703), this is output as the output value of the noise suppression apparatus 100.

<Effect>
In the present embodiment, a noise signal model parameter estimating means for estimating parameters of a multimodal noise signal probability model, and a speaker independent speech signal model for adapting to a speaker dependent speech signal probability model Speaker adaptive parameter estimation means for estimating the parameters is provided, and the parameters are updated while simultaneously executing the parameters. With such a configuration, unsteady noise that is not known in advance can be suppressed with higher accuracy by reflecting speaker characteristics. Moreover, even if the noise signal included in the signal is nonstationary noise that follows a multimodal distribution, the stochastic model parameters (noise model parameters and speaker adaptation parameters) can be easily optimized without solving the nonlinear problem. It is possible to design an optimal noise suppression filter and obtain a target speech signal with high quality.

Note that the estimated noise signal and speech signal may contain errors, but in the estimation of the noise signal probability model and speaker adaptation, the statistical properties of the training data are estimated and processed. Therefore, the problem of error does not become a fatal problem.
<Other variations>
In the first embodiment, each unit and each unit may store each signal, a signal in the middle of processing, various parameters, and the like in a storage unit (not shown), and read / write each data via the storage unit.

In a first embodiment, the frame processing rectangular window besides Hamming window in the window function w _n in (s201 in Fig. 3), a Hanning window may be used a window function, such as Blackman windows.

  In the first embodiment, other probability models such as a Hidden Markov Model (HMM) may be used instead of the silent GMM and the clean speech GMM as the probability model of the sound signal.

  In the first embodiment, not only two GMMs, the silent GMM and the clean voice GMM, but more GMMs may be used. For example, a silent GMM, an unvoiced sound GMM, a voiced sound GMM, or a GMM for each phoneme may be used.

  In the first embodiment, instead of the noise GMM, another probability model such as an HMM may be used as a noise signal probability model.

  In the first embodiment, speaker adaptation processing may be performed using an R × R dimensional affine transformation matrix A as shown in the following equation.

  In the first embodiment, the speaker adaptation process may be performed using an R × R dimensional affine transformation matrix A and a vector b as shown in the following equation.

  In the first embodiment, the affine transformation matrix A, which is a parameter for speaker adaptation processing, and the vector b, the index j for identifying the silence GMM and the clean speech GMM, and the silence GMM or the clean speech as shown in the following equation: It may be a parameter depending on the number k of the normal distribution included in the GMM.

Or

Or

In the first embodiment, in the noise suppression filter estimation process (s603 in FIG. 11), not the weighted average but the maximum weight, that is, the maximum speech non-existence / existence probability P _{t, j} and the posterior probability P _{t, j , K, l} may be used as they are. In this case, it is desirable to have a sufficiently large weight compared to the weights of other estimation results.

  In the first embodiment, instead of the formulas (26), (29), and (30), the initial value may be set by the following formula.

In the first embodiment, the initial value may be set by the following equation instead of the equation (37).

  In 1st embodiment, it is good also as a structure which implements E-step after implementing M-step in each EM algorithm. Moreover, it is good also as a structure which performs convergence determination immediately after implementing M-step, and implements E-step when it has not converged. By setting it as such a structure, the process which implements M-step when it has converged can be skipped.

The present invention is not limited to the above-described embodiments and modifications. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.
<Simulation results>
In order to show the effect of the present invention, an example is shown in which an acoustic signal in which a voice signal and a noise signal are mixed is input to the noise suppression device of the first embodiment and noise suppression is performed. The experimental method and results will be described below.

  In this experiment, the evaluation data uses 100 sentences spoken by 23 men from IPA (Information-technology promotion agency, Japan) -98-TestSet. Noise separately recorded in the lobby, station platform, and street was superimposed on the computer at S / N ratios (signal to noise ratio) of 0 dB, 5 dB, and 10 dB, respectively. That is, a total of nine types of evaluation data of three types of noise × three types of S / N ratios were created. Each audio data is a monaural signal discretely sampled at a sampling frequency of 16,000 Hz and a quantization bit number of 16 bits. The acoustic feature extraction unit 104 was applied to this acoustic signal by setting the time length of one frame to 20 ms (Frame = 320 sample points) and moving the start point of the frame every 10 ms (Shift = 160 sample points).

  As the silent GMM and the clean speech GMM, GMMs having a mixed distribution number K = 128 having an R = 24-dimensional logarithmic mel spectrum as acoustic features are used, and learning is performed using speech data for learning of a large number of speakers. L = 4 was given to the number of mixed distributions of the noise GMM.

  The performance was evaluated by speech recognition, and the evaluation scale was the word error rate (WER) of the following formula.

In the above equation, N is the total number of words, D is the number of dropped error words, S is the number of replacement error words, I is the number of insertion error words, and the smaller the WER value, the higher the speech recognition performance.

  Speech recognition is based on a finite state transducer based recognizer (T. Hori, et al., "Efficient WFST-based one-pass decoding with on-the-fly hypothesis rescoring in extremely large vocabulary continuous speech recognition", IEEE Trans. on ASLP, May 2007, vol. 15, no. 4, pp. 1352-1365), and the speaker model is a speaker-independent Triphone HMM, and each HMM has a three-state Left-to- It is a right type HMM, and each state has 16 normal distributions. The number of states of the entire HMM is 2,000. The acoustic feature of speech recognition is a 12-dimensional MFCC (Mel-frequency cepstral) in which the time length of one frame is 20 ms (Frame = 320) and the start point of the frame is moved every 10 ms (Shift = 160 sample points). coefficient), the logarithmic power value, and a 39-dimensional vector in total including the first and second order regression coefficients. The language model uses Tri-gram and the number of vocabulary is 20,000 words.

FIG. 13 shows the result of noise suppression. Speech recognition according to the case where noise suppression processing is not performed, the method disclosed in Non-Patent Document 1, the method disclosed in Non-Patent Document 2, and the first embodiment. The evaluation results are shown. From the results of FIG. 13, it is clear that higher performance can be obtained by the first embodiment than in the prior art.
<Program and recording medium>
The noise suppression device described above may be realized by causing a computer to execute a program described by a computer-readable code. These programs are stored in a computer-readable storage medium such as a magnetic disk or a CD-ROM, and installed in the computer from the storage medium or installed through a communication line and executed.

  The present invention can be used for automatic speech recognition using a noise suppression signal by suppressing noise from an acoustic signal before the automatic speech recognition. In addition, it can be used when a noise signal is suppressed from a received or recorded sound signal in a call system such as a TV conference system or a recording system.

Claims (7)

A noise suppression device that suppresses a noise signal from an acoustic signal including a noise signal and a voice signal,
Acoustic feature extraction means for extracting acoustic features of the acoustic signal;
Storage means for storing a speaker independent speech model that is a probability model of a speaker independent speech signal learned using speech signals of a large number of speakers without learning as noise.
Speaker adaptation for defining the noise signal as a signal based on non-stationary noise following a multimodal distribution, and adapting the speaker independent speech model to a speaker of the speech signal included in the acoustic signal A first probability model generating means for generating a first probability model that is a probability model of the acoustic signal using a parameter, a noise model that is a probability model of the noise signal, and the speaker independent speech model;
Noise model estimation means for estimating the noise signal based on the first probability model and the acoustic characteristics of the acoustic signal, and unsupervised learning of the noise model using the estimated noise signal as learning data;
Using the acoustic characteristics of the acoustic signal, the speaker independent speech model, and the noise model, the speech signal included in the acoustic signal is estimated, and the estimated speech signal is used as learning data to determine the speaker adaptation parameter. Speaker adaptive parameter estimating means for unsupervised estimation,
Noise suppression means for suppressing a noise signal included in the acoustic signal using an acoustic feature of the acoustic signal, the speaker independent speech model, the noise model, and the speaker adaptation parameter;
Including a noise suppression device. The noise suppression device according to claim 1,
The speaker adaptation parameter estimation means includes:
Second probability model generation means for generating a second probability model that is a probability model of the acoustic signal using the acoustic features of the acoustic signal, the speaker independent speech model, the noise model, and the speaker adaptation parameter; ,
Voice signal estimation means for estimating the voice signal included in the acoustic signal using the acoustic characteristics of the acoustic signal, the second probability model, the speaker independent voice model, and the speaker adaptation parameter;
Based on the speech signal included in the estimated acoustic signal and the speaker independent speech model, the likelihood of the speaker dependent speech model that is the speaker independent speech model adapted by the speaker adaptation parameter is Including speaker adaptation parameter calculation means for calculating the speaker adaptation parameter that is maximized,
Noise suppression device. The noise suppression device according to claim 2,
The speaker adaptation parameter estimation means calculates the speaker adaptation parameter using the estimated speech signal until a convergence condition is satisfied by an expected value maximization method so that the likelihood of the speaker-dependent speech model is maximized. Repeat the process of means,
The noise suppression device uses the acoustic signal and the first probability model generation unit and the noise model estimation until a convergence condition is satisfied by an expected value maximization method so that the likelihood of the first probability model is maximized. Repeating the process of the means and the speaker adaptation parameter estimation means,
Noise suppression device. A noise suppression method for suppressing a noise signal from an acoustic signal including a noise signal and a voice signal,
An acoustic feature extracting means for extracting an acoustic feature of the acoustic signal;
The first probability model generation means defines that the noise signal is a signal based on non-stationary noise that follows a multimodal distribution, and a speech that has been learned using speech signals of many speakers that do not include noise as learning data. A speaker adaptation parameter for adapting a speaker independent speech model which is a probability model of a speech signal independent to a speaker of the speech signal included in the acoustic signal, and a noise model which is a probability model of the noise signal; Generating a first probability model that is a probability model of the acoustic signal using the speaker independent speech model; and
A noise model estimation unit, wherein the noise model estimation means estimates the noise signal based on the first probability model and the acoustic features of the acoustic signal, and performs unsupervised learning of the noise model using the estimated noise signal as learning data. When,
Speaker adaptation parameter estimation means estimates the speech signal included in the acoustic signal using the acoustic features of the acoustic signal, the speaker independent speech model, and the noise model, and learns the estimated speech signal As the data, a speaker adaptation parameter estimation step for estimating the speaker adaptation parameter without teacher,
A noise suppression step, wherein the noise suppression means suppresses a noise signal included in the acoustic signal using an acoustic feature of the acoustic signal, the speaker independent speech model, the noise model, and the speaker adaptation parameter;
Including a noise suppression method. The noise suppression method according to claim 4,
The speaker adaptation parameter estimation step includes:
Second probability model generation means generates a second probability model, which is a probability model of the acoustic signal, using the acoustic feature of the acoustic signal, the speaker independent speech model, the noise model, and the speaker adaptation parameter. A second probability model generation step,
An audio signal for estimating the audio signal included in the audio signal by using an audio feature of the audio signal, the second probability model, the speaker independent audio model, and the speaker adaptation parameter. An estimation step;
A speaker adaptation parameter calculation means is a speaker which is the speaker independent speech model adapted by the speaker adaptation parameter based on the speech signal included in the estimated acoustic signal and the speaker independent speech model. A speaker adaptation parameter calculation step for calculating the speaker adaptation parameter that maximizes the likelihood of the person-dependent speech model,
Noise suppression method. The noise suppression method according to claim 5,
The speaker adaptation parameter estimation step uses the estimated speech signal to calculate the speaker adaptation parameter until a convergence condition is satisfied by an expected value maximization method so that the likelihood of the speaker-dependent speech model is maximized. Repeat the process of steps,
The noise suppression method uses the acoustic signal to generate the first probability model and the noise model estimation until a convergence condition is satisfied by an expected value maximization method so that the likelihood of the first probability model is maximized. Repeating steps and the speaker adaptation parameter estimation step,
Noise suppression method.   The program for functioning a computer as a noise suppression apparatus in any one of Claim 1 to 3.

Images