JP2014029407A - Noise suppression device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a noise suppression technology which suppresses an unsteady noise signal with respect to an input signal.SOLUTION: Input signal spectral information and sound feature quantity used for parameter estimation of a noise model are extracted from an input signal by using a soundless model and a voice model; a noise model parameter of a noise model and a speaker adaptation parameter dependent on a speaker are estimated by using the sound feature quantity, the soundless model and the voice model; a noise suppression filter is estimated by using the sound feature quantity, the soundless model, the voice model, the noise model parameter and the speaker adaptation parameter; and a target signal is output by applying the noise suppression filter to the spectral information. The noise model is expressed as a weighting mixed probability distribution, and a posterior probability used for the estimation of the noise model parameter and the speaker adaptation parameter is estimated by marginalizing the soundless model, the weighting mixed probability distribution of the input signal acquired by synthesizing the sound model and the noise model, and a posterior probability acquired from the input signal.

Description

  The present invention relates to a noise suppression technique for extracting a desired signal by suppressing a noise signal from a signal including a plurality of acoustic signals.

  When the automatic speech recognition technology is used in an actual environment, it is necessary to remove noise from a signal other than the speech signal to be processed, that is, an acoustic signal including noise, and extract only a desired speech signal. 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 Document 1 uses a signal obtained by mixing a speech signal and a noise signal as an input, and generates a probability model of the input signal 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 noise signal constituting the probability model of the input signal and the statistics of each of the speech signal and noise signal included in the input signal is expressed by Taylor series approximation. The difference is estimated using the EM algorithm, and the input signal probability model is optimized. After that, noise is suppressed using the optimized input signal probability model and speech signal probability model parameters.

  Non-Patent Document 2 uses a signal obtained by mixing a voice signal and a noise signal as an input, and a speaker of a voice signal included in the input signal based on a probability model of a voice signal learned using voice data for learning of many speakers. In order to adapt to the characteristics of the speaker (speaker adaptation) and deal with noise signals whose statistical properties follow a multimodal distribution, the speech signal and the noise signal are extracted from the input signal, respectively, Then, the parameters of speaker adaptation using noise signals and the stochastic model of noise signals following multimodal distribution are estimated by the nested EM algorithm. Then, the optimal probability model of the input signal is generated from the probability model of the speech signal after speaker adaptation and the estimated probability model of noise, and the optimal probability model of the input signal and the probability of the speech signal after speaker adaptation Suppresses noise using model parameters.

P. J. Moreno, B. Raj, and R. M. Stern, "A vector Taylor series approach for environment-independent speech recognition," in Proceedings of ICASSP '96, vol. II, pp. 733-736, May 1996. Masayoshi Fujimoto, Tomoharu Watanabe, Tomohiro Nakatani, “Noise Suppression by Simultaneous Application of Speaker Adaptation and Noise Mixing Model Estimation,” IEICE, Speech Society, SP2011-91, pp. 113-118, Dec. 2011.

  A technology essential for automatic speech recognition in an actual environment is a noise suppression technology that removes noise from an input acoustic signal and obtains a high-quality speech signal.

  In Non-Patent Document 1, using the collected input signal, the probability model of each of the speech signal and noise signal constituting the probability model of the input signal is optimized by the EM algorithm, and accordingly the probability model of the input signal is optimized. A technique for noise suppression based on the premise that the characteristics of the noise signal included in the input signal are stationary and the distribution (frequency distribution or probability distribution) is unimodal. It is. However, many noise signals in the real environment have non-stationary characteristics, 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 speech signal and noise signal included in the input signal is expressed by a nonlinear function, even when Taylor series approximation is used, an analytical solution is available when estimating the parameters of the probability model of the speech signal and noise signal. I can't get it. For this reason, the technique described in Non-Patent Document 1 cannot obtain the optimal solution of the probability model parameters of the speech signal and the noise signal, and cannot obtain sufficient noise suppression performance.

  In Non-Patent Document 2, a speech signal and a noise signal are extracted from the collected input signal, and the speaker adaptation parameters and the multimodal distribution are extracted using the extracted speech signal and the noise signal. A method for estimating a stochastic model of a noise signal is disclosed, and it is possible to deal with a non-stationary noise signal. In the technique described in Non-Patent Document 2, a speaker adaptation parameter and a noise signal probability model according to a multimodal distribution are estimated by different EM algorithms. In addition, a probability model of the input signal is generated from the probability model of the speech signal after speaker adaptation and the estimated noise probability model, and the estimation is repeated by the EM algorithm until the parameters of the input signal probability model converge. . For this reason, the technique described in Non-Patent Document 2 has three different EM algorithms, and each has a nested structure, which requires a large amount of calculation and is not a realistic process. . Further, when estimating speaker adaptation parameters and a noise signal probability model according to a multimodal distribution using the speech signal extracted from the input signal and the noise signal, the speech signal, the noise signal, When many extraction errors are included, the speaker adaptation parameters and the estimation accuracy of the noise signal probabilistic model according to the multimodal distribution deteriorate. However, the technique described in Non-Patent Document 2 does not consider such extraction errors.

  In view of such a situation, the present invention uses a noise signal probabilistic model based on a multimodal distribution, and can effectively suppress non-stationary noise signals with respect to an input signal. To provide technology.

  The noise suppression technology of the present invention is a noise suppression technology that uses a sound signal including a speech signal and a noise signal as an input signal and outputs a signal (target signal) in which the noise signal is suppressed from the input signal. Using the signal probability model (silence model) and the sound signal probability model (speech model), from the input signal, the spectrum information of the input signal and the sound used for parameter estimation of the noise signal probability model (noise model) In order to express a noise model parameter (noise model parameter) and a probability model using an acoustic feature quantity extraction process that outputs the feature quantity, an acoustic feature quantity, a silence model, and a speech model A parameter estimation process for estimating a speaker-dependent parameter (speaker adaptation parameter), an acoustic feature amount, a silence model, and a speech module. Performed and Le, the noise model parameters, by using the speaker adaptation parameter, and estimate the noise suppression filter and a noise suppressing process of outputting the target signal by further applying the noise suppression filter to spectral information. The noise model is expressed as a weighted mixed probability distribution with a plurality of probability distributions. In the parameter estimation process, the posterior probabilities used for estimating the noise model parameters and the speaker adaptation parameters are expressed as a silence model, a speech model, and a noise model. Is estimated by marginalizing the weighted mixed probability distribution of the input signal obtained by combining the posterior probabilities obtained from the input signal.

  According to the present invention, as will be described in detail later, it is possible to effectively suppress a non-stationary noise signal with respect to an input signal, and since many EM algorithms are not configured to be nested, many calculations are performed. No need for quantity.

The figure which shows the function structural example of the noise suppression apparatus of embodiment by this invention. The figure which shows the process sequence of an acoustic feature extraction part. The figure which shows the function structural example of a parameter estimation part. The figure which shows the process sequence of a parameter estimation part. The figure which shows the selection method of the reliability data based on SN ratio. The figure which shows the function structural example of a noise suppression part. The figure which shows the process sequence of a noise suppression filter estimation part. The figure which shows the process sequence of a noise suppression filter application part. The figure which shows the selection method of the reliability data based on logarithmic probability ratio. The figure which shows the Example of the noise suppression by this invention.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the drawings used for the following description, the same components are denoted by the same reference numerals. Their names and functions are also the same, and description thereof will not be repeated. In the following explanation, 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 text notation. To do. In the formula, these symbols are written in their original positions. The vector is indicated as A ^{→ in} the text, but is indicated in bold in the formula. The matrix is usually written in font in the text, but in bold in the formula. Further, the processing performed for each element of a vector is applied to all the elements of all vectors and matrices unless otherwise specified.

<Outline>
In order to estimate the parameters of speaker adaptation and the stochastic model of the noise signal according to the multimodal distribution, it is included in the stochastic model of the speech signal and the stochastic model of the noise signal in addition to the speech signal and the noise signal. A posteriori probability (occupancy) for each distribution is required. The posterior probability is defined as the expected value of the hidden variable.The hidden variable is a parameter that indicates from which distribution the signal at a certain time is generated in the probability model, and is a parameter that cannot be observed directly from the signal alone. Because it is, it is called a hidden variable. The posterior probability, which is the expected value of the hidden variable, can be easily estimated by using the EM algorithm. However, the EM algorithm is an iterative parameter estimation algorithm, and the amount of calculation is relatively large. For this reason, when different EM algorithms are applied in estimating the speaker adaptation parameters and the stochastic model of the noise signal according to the multimodal distribution, the amount of calculation is greatly increased. In addition, in order to optimize the probability model of the input signal generated from the probability model of the speech signal after speaker adaptation and the estimated noise probability model, a different EM algorithm is required. The EM algorithm for estimating the stochastic model of the noise signal according to the parameters of speaker adaptation and the multimodal distribution is a nested structure of the EM algorithm for optimizing the stochastic model of the input signal, so a large number of calculations Requires an amount.

  Further, the extraction accuracy of the audio signal and noise signal from the input signal depends on the SN ratio (signal-to-noise ratio) for each unit time. That is, at each time point of the input signal, at locations where the energy of the audio signal is dominant (high SN ratio), the audio signal extraction accuracy is high and the noise signal extraction accuracy is low. Conversely, at locations where the audio signal energy is inferior (low SN ratio), the audio signal extraction accuracy is low and the noise signal extraction accuracy is high. If estimation of a stochastic model of a noise signal according to speaker adaptation parameters and multimodal distribution is performed without considering such extraction accuracy, each estimation performance deteriorates.

  Therefore, in the present invention, the speech signal probability model by the marginalization of the posterior probability of the input signal probability model and the posterior probability estimation of the noise signal probability model, and the extraction accuracy of the sound signal and noise signal from the input signal The estimation of the stochastic model of the noise signal according to the parameters of speaker adaptation and the multimodal distribution in consideration of By using such a method, it is possible to eliminate the nested structure of the EM algorithm and greatly reduce the amount of calculation, and the parameter of the speaker adaptation and the stochastic model of the noise signal according to the multimodal distribution The estimation performance can be improved.

  In the embodiment, a mixture normal distribution (Gaussian mixture model: GMM) is adopted as a probability model of a speech signal and a probability model of a noise signal based on a multimodal distribution.

<Configuration of noise suppression device 100 as a whole>
One embodiment of a noise suppression apparatus according to the present invention is shown in FIG. Reference numeral 100 in the figure indicates the configuration of the noise suppression apparatus according to the embodiment of the present invention. The noise suppression device 100 extracts an acoustic feature extraction that extracts a complex spectrum 112 and a logarithmic mel spectrum 113, which are features for performing noise suppression, from an input signal 103 obtained by mixing a speech signal 101 and a noise signal 102. Unit 104, log mel spectrum 113, silence GMM 108 stored in GMM storage unit 107, clean speech GMM 109, speaker adaptation parameter 110 which is a noise probability model, and noise GMM parameter set Using parameter estimation unit 105 for estimating 111, complex spectrum 112, log mel spectrum 113, silence GMM 108, clean speech GMM 109, speaker adaptation parameter 110, and noise GMM parameter set 111 And a noise suppression unit 106 that designs a noise suppression filter and obtains a noise suppression signal 114 by suppressing noise.

<Configuration of acoustic feature extraction unit 104>
The acoustic feature extraction unit 104 performs processing according to the flow shown in FIG.
First, in frame cut-out processing S201, an input signal o _τ 103 (τ is a sampling point of a discrete signal) is cut out as a frame with an acoustic signal having a fixed time length while moving the start point with a fixed time width in the time axis direction. For example, the acoustic signal o _{t, n} of Frame = 320 sample points (16,000 Hz × 20 ms) is cut out while shifting the start point by Shift = 160 sample points (16,000 Hz × 10 ms). At that time, I cut out for example by multiplying the window function w _n, such as the following Hamming window. Here, t represents the frame number, and n represents the nth sample point in the frame.

After that, by applying fast Fourier transform processing S202 of M points (M is set to a power of 2 and a value equal to or greater than Frame, for example, 512) for o _{t, n} , complex spectrum Spc ^→ _t = {Spc _{t, 0} ,…, Spc _{t, m} ,…, Spc _{t, M−1} } 112 is obtained (m is the frequency bin number). Next, mel filter bank analysis processing S203 and logarithmic processing S204 are applied to the absolute value of Spct _{, m} , and a vector O ^→ _t = {having an R-dimensional (for example, R = 24) log mel spectrum as an element. O _{t, 0} ,…, O _{t, r} ,…, O _{t, R-1} } 113 is calculated (r is the element number of the vector). That is, the output of the acoustic feature extraction unit 104 is a complex spectrum Spc ^→ _t 112 and a log mel spectrum O ^→ _t 113. The complex spectrum Spc ^→ _t 112 is input to the noise suppression unit 106, and the log mel spectrum O ^→ _t 113 is input to the parameter estimation unit 105 and the noise suppression unit 106.

<Outline of parameter estimation unit 105 and parameter definition>
The parameter estimation unit 105 receives the log mel spectrum O ^→ _t 113, the silence GMM 108, and the parameters of the clean speech GMM 109, and optimally estimates the noise GMM parameter set 110 and the speaker adaptation parameter 111.

The silent GMM 108 and the clean voice GMM 109 are given by the following equations.

In the above equation (2), j is an index for identifying the silent GMM 108 and the clean voice GMM 109, j = 0 indicates the silent GMM 108, j = 1 indicates the clean voice GMM 109, and k indicates the silent GMM 108. Alternatively, the normal distribution number included in the clean speech GMM 109, K is the total normal distribution number (for example, K = 128). S ^→ _t and p _{SI, j} (S ^→ _t ) are the log mel spectrum of speech and the likelihood of silence GMM 108 or clean speech GMM 109, and w _{SI, j, k} and , Μ ^→ _{SI, j, k} and Σ _{S, j, k} are the mixing weight, average vector, and diagonal dispersion matrix of the silent GMM 108 or the clean speech GMM 109, respectively. Each parameter is estimated in advance using speech data for learning of many speakers. The function N (•) is a probability density function of a multidimensional normal distribution given by the following equation (3).

A GMM that consists of parameters estimated from the speech data for learning from a large number of speakers is called a speaker independent (SI) GMM, and it consists of parameters estimated from the speech data for learning for a specific speaker. GMM is called speaker dependent (SD) GMM. Since a speaker-dependent GMM is not practical in practice, an SD GMM is obtained by speaker adaptation processing. That is, the average vector μ ^→ _{SD, j, k} of the SD GMM is obtained by converting the average vector μ ^→ _{SI, j, k} of the SI GMM by speaker adaptation processing of the following equation (4).

In the above equation, A ^→ and b ^→ are an R × R dimensional transformation matrix and an R dimensional bias vector, respectively. W and ξ ^→ _{j, k} are an (R + 1) × R-dimensional extended transformation matrix and an R + 1-dimensional extended average vector given by the following equations (5) and (6). The person adaptation parameter 110 is used. W is an independent parameter for the index j for identifying the silent GMM 108 and the clean speech GMM 109, and the normal distribution number k included in the silent GMM 108 or the clean speech GMM 109.

On the other hand, the noise GMM is given by the following equation (7).

In the above equation (7), l is the number of the normal distribution included in the noise GMM, and L is the total normal distribution number (for example, L = 4). N ^→ _t and p _N (N ^→ _t ) are the log mel spectrum of noise and the likelihood of noise GMM, w _{N, l} , μ ^→ _{N, l} and Σ _{N, l} Are a noise GMM mixing weight, an average vector, and a diagonal dispersion matrix, respectively. Hereinafter, the parameter set 111 of the noise GMM is defined as λ = {w _{N, l} , μ _{N, l} , Σ _{N, l} }.

  In the parameter estimation unit 105, the speaker adaptation parameter W 110 and the noise GMM parameter set λ 111 are estimated by the EM algorithm. The EM algorithm is a method used for parameter estimation of a certain probability model. Expectation-step (E-step) that calculates the expected value of the cost function (log-likelihood function) of the probability model and the cost function are maximized. The parameters are optimized by repeating Maximization-step (M-step) until the convergence condition is satisfied.

<Configuration of Parameter Estimator 105>
FIG. 3 shows the configuration of the parameter estimation unit 105, which includes an initialization unit 301, a probability / signal estimation unit 302, a confidence data selection unit 303, a speaker adaptive parameter estimation unit 304, and a noise GMM estimation unit 305. And a convergence determination unit 306. The parameter estimation unit 105 performs processing according to the flow shown in FIG.

<Initialization unit 301>
First, the iteration index i of the EM algorithm is initialized (S401).

Next, in initial value estimation processing S402, initial values of speaker adaptation parameter W 110 and noise GMM parameter set λ 111 in the EM algorithm are estimated by the following equation. U is the number of frames required for initial value estimation (for example, U = 10).

In the above equation, the subscript (i) indicates a parameter in the i-th iterative estimation in the EM algorithm. In Equation (8), 0 ^→ and I are R-dimensional vertical vectors and R × R-dimensional unit matrices, respectively, where all elements are 0, and GaussRand (•) is a multidimensional normal random number generator. .

<Probability / signal estimation unit 302>
In the probabilistic model parameter generation processing S403, the speaker adaptation parameter W ^(i-1) , the noise GMM parameter set λ ^(i-1) , the silence GMM 108, and the clean speech GMM 109 in the ^i- 1th iteration estimation The probability model of the log mel spectrum O ^→ _t 113 is constructed by the following GMM using the parameters

In the above equation (14), p _{O, j} ⁽ⁱ⁾ (O ^→ _t ) is the likelihood of the probability model of the log mel spectrum O ^→ _t 113 generated in the probability model parameter generation processing S403, and w _{O , J, k, l} ⁽ⁱ⁾ , μ ^→ _{O, j, k, l} ⁽ⁱ⁾ and Σ _{O, j, k, l} ⁽ⁱ⁾ are speaker adaptations in the i-1th iteration estimation A mixture of stochastic models of log mel spectrum O ^→ _t 113 generated from parameter W ^(i-1) , noise GMM parameter set λ ^(i-1) , and silence GMM 108 or clean speech GMM 109 parameters The weight, average vector, and diagonal dispersion matrix are given by the following equation.

In the above equation, the functions log (•) and exp (•) perform an operation for each vector element. 1 ^→ is an R-dimensional vertical vector with all elements being 1, and H _{j, k, l} ⁽ⁱ⁾ is a Jacobian matrix of the function h (·).

In the expected value calculation processing S404, the expected value of the cost function Q _O (•) of the probability model of the log mel spectrum O ^→ _t 113 in the i-th iterative estimation is calculated by the following equation (E-step of the EM algorithm).

In the above equation, O ^→ _{0: T-1} = {O ₀ ,…, O _t ,…, O _T-1 }, and T is the total number of frames of the log mel spectrum O ^→ _t 113, P _{t, j} ^{( i)} and P _{t, j, k, l} ⁽ⁱ⁾ are posterior probabilities for GMM type j or normal distribution number k and l in frame t given by the following equations, respectively. 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 existence probability.

  The M-step of the EM algorithm includes a signal estimation process S405, a posteriori probability estimation process S406, a confidence data selection process S407, a speaker adaptive parameter estimation process S408, and a noise GMM parameter estimation process S409.

In the signal estimation process S405, the log mel spectrum S ^→ _t ^{(i) of} the clean speech used to update the speaker adaptation parameter W ⁽ⁱ⁾ and the parameter set λ ^{(i) of} the noise GMM, and the log of the noise The mel spectrum N ^→ _t ⁽ⁱ⁾ is estimated from the log mel spectrum O ^→ _t 113. The log mel spectrum S ^→ _t ^{(i) of} clean speech and the log mel spectrum N ^→ _t ^{(i) of} noise are estimated by the following equations.

In the posterior probability estimation process S406, the posterior probabilities P _{t, j, k} ⁽ⁱ⁾ and P _{t, l} necessary for updating the speaker adaptation parameter W ⁽ⁱ⁾ and the parameter set λ ⁽ⁱ⁾ of the noise GMM ⁽ⁱ⁾ is estimated by the marginalization of the following equation.

<Reliable data selection unit 303>
In the reliable data selection process S407, the estimated log mel spectrum ^ S ^→ _t of the reliable clean speech used when estimating the speaker adaptation parameter W ⁽ⁱ⁾ and the noise GMM parameter set λ ⁽ⁱ⁾ and ^(i), to select the noise of the estimated logarithmic Mel spectrum ^ N ^→ _t ^(i).

First, the SN ratio SNR _t ⁽ⁱ⁾ in frame t is calculated by the following equation.

Next, k-mean clustering is applied to the signal-to-noise ratio SNR _t ⁽ⁱ⁾ obtained from all frames to classify SNR _t ⁽ⁱ⁾ into two classes C = 0, 1, and the average SN of each class The ratio is defined as AveSNR _c ⁽ⁱ⁾ . After that, based on the determination criteria shown in FIG. 5, it is determined which of the clean voice and noise is dominant in the frame t. If the clean voice is dominant, the frame number t is set to a set T _S ⁽ Store in ⁱ⁾ . Conversely, if the noise is dominant, the frame number t is stored in the noise frame set T _N ⁽ⁱ⁾ .

<Speaker adaptive parameter estimation unit 304>
Speaker adaptive parameter estimation processing S408 includes log mel spectrum ^ S ^→ _t ^{(i) of} clean speech estimated in signal estimation processing S405 and posterior probability P _{t, j, k} ⁽ estimated in posterior probability estimation processing S406. ^{Using i)} and the set T _S ⁽ⁱ⁾ of clean speech signal frames estimated in the trust data selection process S407, the speaker adaptation parameter W ⁽ⁱ⁾ is updated by the following equation.

<Noise GMM estimation unit 305>
The noise GMM parameter estimation process S409 includes the log mel spectrum of noise estimated in the signal estimation process S405 ^ N ^→ _t ⁽ⁱ⁾ , the posterior probability P _{t, l} ⁽ⁱ⁾ estimated in the posterior probability estimation process S406, The noise GMM parameter set λ ⁽ⁱ⁾ is updated by the following equation using the set of noise signal frames T _N ⁽ⁱ⁾ estimated in the reliable data selection processing S407.

<Convergence determination unit 306>
In the convergence determination process S410, it is determined whether or not the convergence condition is satisfied. If satisfied, W = W ⁽ⁱ⁾ and λ = λ ⁽ⁱ⁾ are set, and the process of the parameter estimation unit 105 is terminated. If not satisfied, i ← i + 1 (S411), and the process after the probability model generation process S403 is repeated. The convergence condition is the following equation, and η is, for example, η = 0.0001.

<Configuration of Noise Suppression Unit 106>
The configuration of the noise suppression unit 106 is as shown in FIG. 6. The log mel spectrum O ^→ _t 113, the silence GMM 108, the clean speech GMM 109, the speaker adaptation parameter W ^(i), and the noise GMM parameter set In response to λ ⁽ⁱ⁾ , the noise suppression filter estimator 601 for estimating the noise suppression filter F _{t, m} ^Lin , the complex spectrum Spc ^→ _t 112, and the noise suppression filter F _{t, m} ^Lin receives noise. And a noise suppression filter application unit 602 that obtains a noise suppression signal ^ s _τ 114 by suppression.

<Configuration of Noise Suppression Filter Estimation Unit 601>
The noise suppression filter estimation unit 601 performs processing according to the flow shown in FIG.
First, in the probabilistic model generation process S701, the logarithmic mel spectrum O ^→ _t 113 is calculated from the silent GMM 108, the clean speech GMM 109, the speaker adaptation parameter W ^(i), and the noise GMM parameter set λ ^(i). The parameters of the probability model are generated as follows.

In the probability calculation processing S702, and the parameters of the probability model of voice absence / presence probability P _t, and _j, the posterior probability P _{t, j, k,} logarithmic Mel spectrum and _l O ^→ _t 113, logarithmic Mel spectrum O ^→ calculated using the _t 113.

In the noise suppression filter estimation processing S703, the average vector μ ^→ _{SD, j, k of the} SD GMM generated from the extended average vector ξ _{j, k} of the silent GMM 108 and the clean speech GMM 109 and the speaker adaptation parameter W Mel frequency using noise GMM mean vector μ ^→ _{N, l} , speech absence / presence probability P _{t, j} and posterior probability P _{t, j, k, l} included in noise GMM parameter set λ The noise suppression filter F _{t, r} ^Mel on the axis is estimated as follows: Expression (36) is a notation for each element of the vector.

In noise suppression filter estimation processing S704, the noise suppression filters F _{t and r} ^Mel on the mel frequency axis are converted into noise suppression filters F _{t and m} ^Lin on the linear frequency axis. The transformation estimates the value of the noise suppression filter on the linear frequency axis by applying cubic spline interpolation to the mel frequency axis.

<Configuration of Noise Suppression Filter Application Unit 602>
The noise suppression filter application unit 602 performs processing according to the flow shown in FIG.
In the filtering process S801, the complex spectrum Sp _t ^→ _m 112 is multiplied by a noise suppression filter F _{t, m} ^Lin as shown in the following equation to obtain a noise-suppressed complex spectrum ^ S _{t, m} . Expression (37) is a notation for each element of the vector.

In the inverse fast Fourier change processing S802, the noise suppression speech ^ _{st, n} in the frame t is obtained by applying the inverse fast Fourier transform to the complex spectrum ^ St _{, m} .

Obtaining a noise reduced speech ^ s _t, the noise reduced speech ^ s _tau 114 where the _n consecutive linked while releasing the window function w _n as in the following equation for each frame in the waveform linking process S803.

  The noise suppression apparatus 100 is 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 CD-ROM, and installed in the computer from the storage medium or installed through a communication line and executed.

<Modification>
In the above embodiment, the square windows besides Hamming window in the window function w _n at frame cutout processing S201, Hanning window may be used a window function, such as Blackman windows.

  In the above embodiment, instead of the silent GMM 108 and the clean speech GMM 109, another probability model such as a hidden Markov model (HMM) may be used as the speech signal probability model.

  In the above embodiment, not only two GMMs, the silent GMM 108 and the clean voice GMM 109, but a larger number of 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 above embodiment, instead of the noise GMM, another probability model such as an HMM may be used as a noise signal probability model.

In the above embodiment, the speaker adaptation process may be performed using only the transformation matrix A as shown in the following equation.

In the above embodiment, the speaker adaptation process may be performed using only the bias vector b ^→ as in the following equation.

In the above embodiment, conversion matrix A and bias vector b ^→ which are parameters of speaker adaptation processing, an index j for identifying silent GMM 108 and clean speech GMM 109, and silent GMM 108, Alternatively, the parameter may depend on the normal distribution number k included in the clean speech GMM 109.

In the above embodiment, the selection of the reliable data may be performed using the log probability ratio LPR _t ⁽ⁱ⁾ of the following equation instead of the SN ratio SNR _t ⁽ⁱ⁾ .

At this time, k-mean clustering is applied to the log probability ratio LPR _t ⁽ⁱ⁾ obtained from all the frames, and LPR _t ⁽ⁱ⁾ is classified into two classes C = 0, 1, and the average of each class The log probability ratio is defined as AveLPR _c ⁽ⁱ⁾ . Thereafter, it is determined whether the clean voice or the noise is dominant in the frame t according to the criterion shown in FIG. 9. If the clean voice is dominant, the frame number t is set to the set T _S ⁽ Store in ⁱ⁾ . Conversely, if the noise is dominant, the frame number t is stored in the noise frame set T _N ⁽ⁱ⁾ .

In the above embodiment, trust data selection may be performed using threshold processing of the following equation instead of k-mean clustering. Th _SNR and Th _LPR are thresholds corresponding to SNR _t ⁽ⁱ⁾ and LPR _t ⁽ⁱ⁾ , respectively.

In the above embodiment, in the noise suppression filter estimation processing S703, 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} The estimation result having the product of may be used as it is. In this case, it is desirable to have a sufficiently large weight compared to the weights of other estimation results.

  In order to show the effect of the present invention, an example in which an acoustic signal in which a voice signal and a noise signal are mixed is input to the noise suppression device according to the embodiment of the present invention and noise suppression is performed will be shown. The experimental method and results will be described below.

  In this experiment, IPA (Information-technology promotion agency, Japan) -98-TestSet of 100 data uttered by 23 men was used for the evaluation data. Noises recorded separately in the lobby, station platform, and street were superimposed on the computer with SN ratios 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 SN ratio were created. Each audio data is a monaural signal that is 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). .

  For the silent GMM 108 and clean speech GMM 109, R = 24-dimensional log mel spectrum was used as a GMM with a mixed distribution number K = 128, which was trained using speech data for learning of many speakers. . L = 4 was given to the number of mixed distributions of noise GMM.

The performance was evaluated by speech recognition, and the evaluation scale was the word error rate (WER) of the following formula. 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, vol.15, no.4, pp.1352-1365, May 2007), and the acoustic model uses 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 total number of HMMs is 2,000. The acoustic feature for speech recognition is a 12-dimensional MFCC (Mel-) analyzed by moving the start point of the frame every 10 ms (Shift = 160 sample points), with the time length of one frame being 20 ms (Frame = 320). frequency cepstral 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 vocabulary is 20,000 words.

  FIG. 10 shows the results of speech recognition evaluation by the noise suppression result, no noise suppression, the method disclosed in Non-Patent Document 1, the method disclosed in Non-Patent Document 2, and the noise suppression according to the present invention. Show. From the results shown in FIG. 10, it has been clarified that the present invention can obtain higher performance than the prior art.

Claims (5)

A noise suppression device that outputs an acoustic signal including an audio signal and a noise signal as an input signal, and outputs a signal (hereinafter referred to as a target signal) in which the noise signal is suppressed from the input signal,
A storage unit for storing a probability model of a silence signal (hereinafter referred to as a silence model) and a probability model of a sound signal (hereinafter referred to as a speech model);
An acoustic feature quantity extraction unit that outputs spectrum information of the input signal and an acoustic feature quantity used for parameter estimation of a stochastic model of a noise signal (hereinafter referred to as a noise model) from the input signal;
It is used to express a parameter expressing the noise model (hereinafter referred to as a noise model parameter) and a probability model of the acoustic feature using the acoustic feature, the silence model, and the speech model. A parameter estimator for estimating a speaker-dependent parameter (hereinafter referred to as a speaker adaptation parameter);
Estimating a noise suppression filter using the acoustic feature, the silence model, the speech model, the noise model parameter, and the speaker adaptation parameter, and applying the noise suppression filter to the spectrum information And a noise suppression unit that outputs the target signal by
The noise model is expressed as a weighted mixed probability distribution with a plurality of probability distributions,
The parameter estimation unit
The posterior probability used for the estimation of the noise model parameter and the speaker adaptation parameter is obtained from the silence model, the weighted mixed probability distribution of the input signal obtained by combining the speech model and the noise model, and the input signal. A noise suppressor characterized by estimating posterior probabilities by marginalizing. The noise suppression device according to claim 1,
The parameter estimation unit
Estimating the noise model parameter and the speaker adaptation parameter by repeating the calculation process of the expected value of the cost function of the probability model of the acoustic feature quantity and the maximization process of the cost function until a convergence condition is satisfied. A noise suppression device characterized by the above. The noise suppression device according to claim 2,
The parameter estimation unit
In the cost function maximization process, the acoustic feature quantity related to the noise signal and the acoustic feature quantity related to the audio signal are estimated, and furthermore, a reliable one is extracted for each, and the highly reliable The speaker adaptation parameter is updated using the acoustic feature quantity and the posterior probability related to the speech signal, and the noise model parameter is updated using the acoustic feature quantity and the posterior probability related to the highly reliable noise signal. The noise suppression device characterized by updating. An acoustic signal including an audio signal and a noise signal is used as an input signal, and a noise suppression method for outputting a signal (hereinafter referred to as a target signal) in which the noise signal is suppressed from the input signal,
Using a silence signal probability model (hereinafter referred to as a silence model) and a sound signal probability model (hereinafter referred to as a speech model),
An acoustic feature quantity extraction step for outputting, from the input signal, spectral information of the input signal and an acoustic feature quantity used for parameter estimation of a stochastic model of a noise signal (hereinafter referred to as a noise model);
It is used to express a parameter expressing the noise model (hereinafter referred to as a noise model parameter) and a probability model of the acoustic feature using the acoustic feature, the silence model, and the speech model. A parameter estimation step for estimating a speaker-dependent parameter (hereinafter referred to as a speaker adaptation parameter);
Estimating a noise suppression filter using the acoustic feature, the silence model, the speech model, the noise model parameter, and the speaker adaptation parameter, and applying the noise suppression filter to the spectrum information A noise suppression step for outputting the target signal by
The noise model is expressed as a weighted mixed probability distribution with a plurality of probability distributions,
In the parameter estimation step,
The posterior probability used for the estimation of the noise model parameter and the speaker adaptation parameter is obtained from the silence model, the weighted mixed probability distribution of the input signal obtained by combining the speech model and the noise model, and the input signal. A noise suppression method characterized by estimating posterior probabilities by marginalizing.       The program for functioning a computer as a noise suppression apparatus in any one of Claims 1-3.

Images