JP2013007975A - Noise suppression device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a noise suppression technology which can reflect changes and features of more noise signals to a probability model of a noise signal and correctly suppress a noise signal by using a noise signal as learning data, regardless of existence of a voice signal.SOLUTION: A sound feature of a sound signal is extracted. A noise signal is estimated by using a probability model of a voice signal excluding a noise (hereafter, referred to as "voice model") and the sound feature of the sound signal, and an unsupervised learning of a probability model of a noise signal (hereafter, referred to as "noise model") is performed with the estimated noise signal as learning data. A noise signal of the sound signal is suppressed by using the noise model.

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 a noise signal are known in order to make it easier to hear a sound signal from a sound signal including a sound signal to be processed and a signal other than the sound signal (hereinafter referred to as “noise signal”). 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 Document 1 is known as a prior art related to noise suppression. Non-Patent Document 1 generates a probability model of an acoustic signal from a speech signal and a noise signal estimated in advance, and expresses the difference between the probability model and the statistic of the entire acoustic signal by Taylor expansion. The difference is estimated using an EM algorithm (hereinafter also referred to as “expected value maximization method”), and the acoustic signal probability model is optimized. Thereafter, a method of suppressing noise using the optimized parameters of the acoustic signal probability model and the speech signal probability model is disclosed.

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

  In order to estimate the probability model of a noise signal in the prior art, learning data of only the noise signal is required. However, normally, the signal that can be observed when performing noise suppression is only an acoustic signal in which the noise signal and the audio signal are mixed, and it is difficult to observe only the noise signal alone. For this reason, in the prior art, learning data of only the noise signal is obtained by estimating a time interval in which there is no audio signal and only the noise signal exists. However, in such a method, a noise signal in a time interval in which an audio signal exists cannot be used as learning data, and changes and characteristics of the noise signal generated in the interval are reflected in the noise signal probability model. I can't. For this reason, it is difficult to accurately estimate and represent the distribution of the noise signal.

  The present invention can use a noise signal as learning data regardless of the presence or absence of a speech signal, and can reflect more changes and features of the noise signal in the probability model of the noise signal. An object of the present invention is to provide a noise suppression technique that can be suppressed.

  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. A noise signal is estimated using a noise signal-free stochastic model (hereinafter referred to as “speech model”) and the acoustic features of the acoustic signal, and the estimated noise signal is used as training data to determine the noise signal probability model (hereinafter “ "Noise model"). The noise signal of the acoustic signal is suppressed using the noise model.

  The noise suppression technique according to the present invention has an effect that noise signals can be more accurately suppressed.

2 is a functional block diagram of the noise suppression device 100. FIG. The figure which shows the processing flow of the noise suppression apparatus. The figure which shows the processing flow of the acoustic feature extraction part 104. FIG. The functional block diagram of the noise model estimation part 105. FIG. The figure which shows the processing flow of the noise model estimation part 105. FIG. The functional block diagram of the noise model parameter estimation means 306. The figure which shows the processing flow of the noise model parameter estimation means 306. FIG. 3 is a functional block diagram of the noise suppression unit 106. 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 example of an estimation of the noise model by this invention. The figure which shows the noise suppression example by this invention. The figure which shows the noise suppression signal with respect to the audio | voice signal contained in the acoustic signal by this invention.

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

<Noise Suppression Device 100 according to First Embodiment>
A noise suppression device 100 according to the first embodiment will be described with reference to FIGS. 1 and 2.

As shown in FIG. 1, the noise suppression apparatus 100 includes an acoustic feature extraction unit 104, a GMM storage unit 107 in which a silent GMM (Gaussian mixture model) and a clean speech GMM constituting a speech model are stored, A noise model estimation unit 105 and a noise suppression unit 106 are included. The noise suppression apparatus 100 is recorded an acoustic signal o _tau where the audio signal and the noise signal are mixed, or, and outputs a noise suppressed signal ^ s _tau was suppressing noise signals from the acoustic signal o _tau. However, (tau) represents the sample point of a discrete signal. Hereinafter, an outline of the present embodiment will be described.

  As shown in FIG. 2, the acoustic feature extraction unit 104 extracts a complex spectrum and a log mel spectrum for performing noise suppression from the acoustic signal (s104). The noise model estimation unit 105 uses a logarithmic mel spectrum and a silence GMM and a clean speech GMM held in the main memory by the GMM storage unit 107 to obtain a noise GMM which is a probability model (hereinafter referred to as “noise model”) of a noise signal. Estimate (s105). The noise suppression unit 106 designs a noise suppression filter using a complex number spectrum, logarithmic mel spectrum, silence GMM, clean speech GMM, and noise GMM, and suppresses the noise signal from the acoustic signal to generate the noise suppression signal. Obtain (s106). Details of each part will be described below.

<Acoustic Feature Extraction Unit 104>
The acoustic feature extraction unit 104 extracts the acoustic feature of the acoustic signal (s104). 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 _o t ₌ frame width Frame = 320 samples points _{(16,000Hz × 20ms) {o t} , 0, o t, 1, ..., o t, n, ..., o t, 319} a, 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. Acoustic signals in frame units and o _t, expressed as follows.
o _t = {o _{t, 0} , o _{t, 1} ,…, o _{t, n} ,…, o _{t, Frame-1} }
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.

Then, fast Fourier point M to the acoustic feature extraction unit 104 the audio signal _{o t} (where, M is a power of 2, and must be set to a minimum of the frame width Frame, eg, 512) Applying the transformation process, a complex spectrum Spc _t = {Spc _{t, 0} ,..., Spc _{t, m} ,..., Spc _{t, M−1} } (where m is the frequency bin number) is obtained (s202 ).

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 Spc _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 noise model estimating section 105, the input of the noise suppressor 106.

<GMM storage unit 107>
In a storage unit (not shown), a probability model (hereinafter referred to as “speech model”) of a speech signal that does not contain noise is stored in advance. For example, a silent GMM and a clean speech GMM are stored as speech models in the GMM storage unit 107 that is a part of the storage unit. The silent GMM is a GMM learned based on an acoustic signal acquired from a silent portion of a speech signal that does not include a noise signal, and the clean speech GMM is based on an acoustic signal consisting only of speech excluding the silent portion in an environment without noise. It is a learned GMM.

  The silent GMM and the clean speech GMM are given by the following equations.

In the above equation, 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, _St is a logarithmic mel spectrum of a speech signal not including noise, and b _{S, j} (S _t ) is a likelihood of a silent GMM or a clean speech GMM. The subscript S indicates a likelihood or parameter relating to an acoustic model (silent GMM or clean speech GMM) different from a noise GMM, which will be described later, or a GMM of a speech signal and an acoustic signal including the noise signal. Yes. Further, w _{S, j, k} , μ _{S, j, k} and Σ _{S, j, k} are a mixing weight of a silent GMM or a clean speech GMM, an average vector, and a diagonal dispersion matrix, respectively. The function N (•) is a probability density function of a multidimensional normal distribution given by the following equation.

  On the other hand, a noise signal GMM (hereinafter referred to as “noise GMM”) can be used as the noise model. The noise GMM is given by:

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 the logarithmic mel spectrum of noise, b _N (N _t ) is the likelihood of noise GMM, and w _{N, l} , μ _{N, l} and Σ _{N, l} are noise respectively. GMM mixture weight, average vector, and diagonal dispersion matrix. Hereinafter, a noise GMM parameter set (hereinafter also referred to as “noise model parameters”) is defined as λ = {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. Note that the noise model estimation unit 105 described later performs unsupervised learning of the noise model.

<Noise Model Estimation Unit 105>
The noise model estimation unit 105 estimates a noise signal by using the log mel spectrum O _t , the silence GMM, and the clean speech GMM, and performs unsupervised learning of the noise GMM using the estimated noise signal as learning data (s105). In the present embodiment, not the noise signal itself but an acoustic feature (log mel spectrum) of the noise signal is estimated, and the noise GMM is learned using this.

  For example, in the noise model estimation unit 105, the noise GMM is estimated by two types of nested EM algorithms. Hereinafter, these two types of EM algorithms will be referred to as a first EM algorithm and a second EM algorithm, respectively. The EM algorithm is a method used for parameter estimation of a certain probability model. Expectation-step (E-step) for calculating an expected value of the cost function (log likelihood function) of the probability model and the cost function are maximized. The parameter is optimally estimated by repeating Maximization-step (M-step) to satisfy the convergence condition.

  In the first EM algorithm, until the convergence condition is satisfied so that the likelihood of the stochastic model of the acoustic signal including the noise signal and the voice signal (hereinafter also referred to as “acoustic model”) is maximized using the acoustic signal. A probability model generation process (s303), a first expected value calculation process (s304), a noise signal estimation process (s305), and a noise model parameter estimation process (s306) described later are repeated (see FIG. 5).

  The second EM algorithm is executed by a noise model parameter estimation unit 306 described later, and a second expected value calculation described later is performed using the estimated noise signal until the convergence condition is satisfied so that the likelihood of the noise GMM is maximized. The process (s403) and the noise GMM parameter update process (s404) are repeated (see FIG. 7).

  Details of the noise model estimation unit 105 will be described below with reference to FIGS. 4 and 5.

  For example, as shown in FIG. 4, the noise model estimation unit 105 includes a first initial value estimation unit 302, a probability model generation unit 303, a first expected value calculation unit 304, a noise signal estimation unit 305, a noise model parameter estimation unit 306, A convergence determination means 307.

(First initial value estimation means 302)
First, the first initial value estimating means 302 initializes an index i indicating the number of repetitions of the first EM algorithm (s301). Next, the first initial value estimation means 302 receives the log mel spectrum O _t, and the initial value λ ^{(i = 0)} = {w ^{(i = 0)} _{N, l} , w ^{( i = 0)} _{N, l} , w ^{(i = 0)} _{N, l} } is estimated by the following equation (s302) and output to the probability model generation means 303. However, A is the number of frames required for initial value estimation (for example, A = 10).

In the above equation, the subscript (i) indicates a parameter in the i-th iteration estimation. Here, diag represents a diagonal matrix with elements in parentheses, and the superscript T represents transposition.

(Probability model generation means 303)
The probability model generation unit 303 generates an acoustic model using the noise GMM, the clean speech GMM, and the silence GMM (s303). For example, the probability model generation unit 303 receives the noise model parameter λ ⁽ⁱ⁾ in the i-th iterative estimation from the first initial value estimation unit 302 or the first convergence determination unit 307 and receives the parameters (w of silent GMM and clean speech GMM). _{S, j, k} , μ _{S, j, k} , Σ _{S, j, k} ) are received from the GMM storage unit 107, and using these values, a probabilistic model of the log mel spectrum O _t is represented by the following GMM. Consists of.

In the above equation, b _{O, j} ⁽ⁱ⁾ (O _t ) is the likelihood of the probability model (of the log mel spectrum O _t ) generated by the probability model generation means 303, and w _{O, j, k, l} ⁽ⁱ⁾ and μ _{O, j, k, l} ⁽ⁱ⁾ and Σ _{O, j, k, l} ⁽ⁱ⁾ are the noise model parameters λ ⁽ⁱ⁾ = {w ⁽ⁱ⁾ _{N, l} , μ ^{( i)} _{N, l} , Σ ⁽ⁱ⁾ _{N, l} } and silent GMM or clean speech GMM parameters (w _{S, j, k} , μ _{S, j, k} , Σ _{S, j, k} ) The mixture weight of the probability model of the log mel spectrum O _t , the average vector, and the diagonal dispersion matrix are given by the following equations.

In the above equation, the functions log (•) and exp (•) perform an operation for each vector element. 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 of the formula (10) (·). Note that the subscript O indicates the likelihood or parameter related to GMM of an acoustic signal including a speech signal and a noise signal. The probabilistic model generation means 303 is the acoustic model parameters w O, j, k, l (i) , μ O, j, k, l (i) and Σ O which are parameters of the acoustic model obtained by the equations (9) to (12). , J, k, l (i) are output to the first expected value calculation means 304.

(First expected value calculation means 304)
The first expected value calculation means 304 is the acoustic model parameters w _{O, j, k, l} ⁽ⁱ⁾ , μ _{O, j, k, l} ⁽ⁱ⁾ and ΣO _{, j, k, l} ^(i). Is received from the probability model generation means 303, the log mel spectrum O _t of the acoustic signal is received from the acoustic feature extraction unit 104, and the expectation of the cost function Q ₁ () of the probability model of the log mel spectrum O _t in the i-th iterative estimation. The value is calculated by the following equation (E-step) (s304).

In the above equation, O _{0: T-1} = {O ₀ ,..., O _t ,..., O _T-1 }, where T is the total number of frames of the log mel spectrum O _t , P ⁽ⁱ⁾ _{t, j} and P ⁽ⁱ⁾ _{t, j, k, and l} 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 ⁽ⁱ⁾ _{t, j = 0} is defined as a speech non-existence probability, and P ⁽ⁱ⁾ _{t, j = 1} is defined as a speech existence probability.

第一期待値計算手段３０４は、求めた第一期待値Ｑ_１を第一収束判定手段３０７に、Ｐ^（ｉ） _ｔ，ｊとＰ^（ｉ） _{ｔ，ｊ，ｋ，ｌ}とを雑音信号推定手段３０５に出力する。

First expectation value calculation unit 304, a first expected value _{Q 1} obtained in the first convergence determining unit _{^{307, P (i) t,}} j and _{^{P (i) t, j,}} k, the noise signal estimate and _l Output to the means 305.

(Noise signal estimation means 305)
The noise signal estimation means 305 estimates a noise signal using the acoustic signal (s305). For example, the noise signal estimation unit 305 receives P ⁽ⁱ⁾ _{t, j} and P ⁽ⁱ⁾ _{t, j, k, l} from the first expected value calculation unit 304 and determines the logarithmic mel spectrum O _t of the acoustic signal as an acoustic feature. The log mel spectrum N ⁽ⁱ⁾ _t of the noise signal received from the extraction unit 104 and used to update the noise model parameter λ ⁽ⁱ⁾ is estimated and output to the noise model parameter estimation means 306. The log mel spectrum N ⁽ⁱ⁾ _t of noise is estimated by the following equation.

(Noise model parameter estimation means 306)
The noise model parameter estimation means 306 estimates the noise model parameter λ ⁽ⁱ⁾ (M-step) (s306) using the logarithmic mel spectrum N ⁽ⁱ⁾ _t of the noise signal as learning data, and the estimated noise model parameter The result is output to the convergence determination means 307. A specific estimation method of the noise model parameter λ ⁽ⁱ⁾ will be described later.

(First convergence determination means 307)
First convergence determining means 307, the first expected value calculation unit 304 receives the first expected value Q _1, to determine the convergence condition is satisfied whether using this value (s307), if it meets lambda = lambda ^(I) is output as λ, and the process of the noise model estimation unit 105 is terminated. If not satisfied, λ ⁽ⁱ⁾ is output to the probability model generation means 303, and as i ← i + 1 (s308), a control signal is output to each unit so as to perform repetition processing, and the processing of s303 to s306 is repeated. For example, the convergence condition includes the latest first expected value Q ₁ (O _{0: T−1} , λ ⁽ⁱ⁾ ) and the previous first expected value Q ₁ (O _{0: T−1} , λ ^{(i−). 1) The} case where the difference from) is less than or equal to a predetermined value, or the number of repetitions i is greater than or equal to a predetermined value. For example, it can be expressed by the following formula.

であり、η_１＝０．０００１とする。

And η ₁ = 0.0001.

<Details of Noise Model Parameter Estimation Unit 306>
The noise model parameter estimation unit 306 includes, for example, a second initial value estimation unit 402, a second expected value calculation unit 403, a parameter update unit 404, and a second convergence determination unit 405 as shown in FIG. The noise model parameter estimation unit 306 performs processing according to the processing flow shown in FIG. 7, and uses the noise log mel spectrum N ⁽ⁱ⁾ _t estimated by the noise signal estimation unit 305 and the noise model parameter λ ^{( i)} is estimated. Details of the noise model parameter estimation means 306 will be described below with reference to FIGS.

(Second initial value estimating means 402)
The second initial value estimating means 402 first initializes an index ii indicating the number of repetitions of the second EM algorithm (s401). Next, the second initial value estimation means 402 receives the log mel spectrum N ⁽ⁱ⁾ _t of noise, and uses this value to determine the initial value λ ^{(ii = 0 )} of the noise model parameter λ ⁽ⁱⁱ⁾ in the second EM algorithm. ⁾ = {w ^{(ii = 0)} _{N, l} , μ ^{(ii = 0)} _{N, l} , Σ ^{(ii = 0)} _{N, l} } is estimated by the following equation and output to the second expected value calculation means 403 .

In the above equation, the subscript (ii) indicates a parameter in the ii-th iteration estimation. GaussRand (·) is a normal random number generator.

(Second expected value calculation means 403)
The second expected value calculation means 403 receives the log mel spectrum N ⁽ⁱ⁾ _t of noise from the noise signal estimation means 305, and receives the noise model parameter λ ⁽ⁱⁱ⁾ in the second EM algorithm as the second initial value estimation means 402 or the first. The expected value of the cost function Q ₂ () of the noise GMM in the ii-th iterative estimation is calculated from the following equation (E-step) (s 403) and output to the second convergence determination unit 405. .

In the above equation, N ⁽ⁱ⁾ _{0: T-1} = {N ⁽ⁱ⁾ ₀ , ..., N ⁽ⁱ⁾ _t , ..., N ⁽ⁱ⁾ _T-1 }, and P ⁽ⁱⁱ⁾ _{t, l} 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 403 outputs the calculated P ⁽ⁱⁱ⁾ _{t, l} to the parameter update unit 404.

(Parameter update means 404)
The parameter updating unit 404 receives P ⁽ⁱⁱ⁾ _{t, l} , receives the log mel spectrum N ⁽ⁱ⁾ _t of noise from the noise signal estimation unit 305, and updates the noise model parameter λ ⁽ⁱⁱ⁾ by the following equation (M -Step) (s404), and outputs the updated noise model parameter λ ⁽ⁱⁱ⁾ to the second convergence determination means 405.

(Second convergence determination means 405)
Second convergence determining means 405, from the second expected value calculation unit 403 receives the second expected value Q _2, to determine the convergence condition is satisfied whether using this value (s405), if it meets lambda ^{(i )} = Λ ⁽ⁱⁱ⁾ , λ ⁽ⁱ⁾ is output, and the processing of the noise model parameter estimation means 306 is terminated. If not satisfied, λ ⁽ⁱⁱ⁾ is output to the second expected value calculation means 403, and as ii ← ii + 1 (s406), a control signal is output to each unit so as to perform repetition processing, and the processing of s403 and s404 is repeated. For example, the convergence condition includes the latest second expected value Q ₂ (O _{0: T−1} , λ ⁽ⁱ⁾ ) and the previous second expected value Q ₂ (O _{0: T−1} , λ ^{(i−). 1) The} case where the difference from) is less than or equal to a predetermined value, or the case where the number of repetitions ii is greater than or equal to a predetermined value. For example, it can be expressed by the following formula.

And η ₂ = 0.0001.

<Noise Suppression Unit 106>
The noise suppression unit 106 suppresses the noise signal of the acoustic signal using the noise GMM (s106). For example, as shown in FIG. 8, the noise suppression unit 106 includes a noise suppression filter estimation unit 501 and a noise suppression filter application unit 502. The noise suppression filter estimation means 501 includes a log mel spectrum O _t of an acoustic signal, parameters {w _{S, j, k} , μ _{S, j, k} , Σ _{S, j, k} } of silence GMM and clean speech GMM, noise The model parameter λ is received and the noise suppression filter W ^Lin _{t, m} is estimated. The noise suppression filter application unit 502 receives the complex spectrum Spc _t and the noise suppression filter W ^Lin _{t, m} , suppresses noise, and obtains a noise suppression signal ^ _sτ . Details of each means will be described below.

(Noise suppression filter estimation means 501)
The noise suppression filter estimation means 501 performs processing according to the flow shown in FIG. First, the noise suppression filter estimation means 501 receives the parameters of the silent GMM and the clean speech GMM and the noise model parameters, and uses these values to set the parameters of the probability model of the log mel spectrum O _t of the acoustic signal as follows. (S601).

Next, the noise suppression filter estimation means 501 uses the probability model parameter of the log mel spectrum O _t obtained from the speech non-existence / presence probability P ⁽ⁱ⁾ _{t, j} and the posterior probability P _{t, j, k, l.} And the log mel spectrum O _t (s602).

Next, the noise suppression filter estimation means 501 includes a silence GMM parameter, a clean speech GMM parameter, a noise model parameter, a speech absence / existence probability P ⁽ⁱ⁾ _{t, j,} and a posteriori probability P _{t, j, k, l.} Is used to estimate the noise suppression filter W ^Mel _{t, r} on the mel frequency axis as in the following equation (s603).

The above equation is a notation for each vector element.

Next, the noise suppression filter estimation means 501 converts the noise suppression filter W ^Mel _{t, r} on the mel frequency axis into a 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 502. 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 502)
The noise suppression filter application unit 502 performs processing according to the flow shown in FIG. Noise suppression filter application unit 502, the noise suppression filter estimator 501 from the noise suppression filter ^W _{Lin t,} receives the _m, the noise suppression filter ^W _{Lin t} for complex spectrum Spc _t received from the acoustic feature extraction unit _104, the _m By multiplying as in the following equation, a noise-suppressed complex spectrum {circumflex over (S)} _{t, m} is obtained (s701).

The above equation is a notation for each vector element.

Next, the noise suppression filter application unit 502 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 502 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>
With such a configuration, not only the noise signal in the time interval in which only the noise signal exists, but also the noise signal in the time interval in which both the noise signal and the audio signal exist can be used as learning data. In other words, a noise signal can be used as learning data regardless of the presence or absence of an audio signal. As a result, more changes and features of the noise signal can be reflected in the probability model of the noise signal, and the noise signal can be modeled more accurately and noise suppression can be performed with high accuracy. Note that the estimated noise signal may contain errors, but in the estimation of the probabilistic model, since the modeling is performed by estimating the statistical properties of the learning data, the error problem is fatal. It does not become a problem.

  Further, Non-Patent Document 1 discloses a method for optimizing a probability model of an acoustic signal by using an EM algorithm using the entire collected acoustic signal, but the characteristics of the noise signal included in the acoustic signal are steady. This is a technique for performing noise suppression on the assumption that the distribution (frequency distribution or probability distribution) is unimodal. 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 a noise signal having unsteady characteristics and having a multimodal distribution, and may not provide sufficient noise suppression performance.

  In the present embodiment, based on the premise that the distribution of the noise signal is not unimodal but multimodal, the noise signal probability model is expressed by GMM instead of a single normal distribution, and the noise signal Are estimated by the EM algorithm. With such a configuration, it is possible to appropriately model a noise signal having non-stationary characteristics and having a multimodal distribution, and the noise signal can be effectively suppressed. Note that if you attempt to model a noise signal that has non-stationary features and its distribution is multimodal, the model will be more complicated than if you are trying to model a noise signal that is unimodal. Although the amount of data increases, as described above, in this embodiment, a noise signal can be used as learning data regardless of the presence or absence of an audio signal, so that a large amount of data can be acquired, and an optimal noise suppression filter Can be designed.

  Therefore, with the configuration as in the present embodiment, even in an environment where various types of noise exist, the target audio signal can be extracted with high quality by suppressing the noise signal from the acoustic signal.

[simulation result]
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 apparatus of the present invention and noise suppression is performed. The experimental method and results will be described below.

  In this experiment, the data for evaluation uses 100 sentences spoken by 23 men from IPA (Information-technology promotion agency, Japan) -98-TestSet. Noises recorded separately in the lobby, station platform, and street were superimposed on the computer with S / N ratios of 0 dB, 5 dB, and 10 dB, respectively. That is, 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, a GMM having a mixed distribution number K = 128 having an R = 24-dimensional logarithmic mel spectrum as an acoustic feature is used, and learning is performed using the silent signal and the clean speech signal, respectively. The mixed distribution number L of the noise GMM gives four values of 1, 2, 3, and 4, and the results in each case are compared.

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

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). Speaker-independent Triphone HMM is used for the acoustic model, and each HMM has a three-state Left-to-right structure. 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. 11 is an example of noise model estimation. Reference numerals 801 to 805 denote noise model estimation results according to the method disclosed in Non-Patent Document 1 and the present invention (L = 1, 2, 3, 4), respectively. The logarithmic mel spectrum distribution of the noise obtained from the eighth mel filter (center frequency: 1022 Hz) is shown. In each figure, the broken line indicates the histogram of the log mel spectrum of noise, the solid line indicates the shape of the noise probability model estimated by each method, and the dotted line indicates the shape of each element distribution constituting the noise GMM. The vertical axis indicates the normalized frequency or probability, and the horizontal axis indicates the value of the noise mel spectrum obtained from the eighth mel filter.

  From the result of FIG. 11, it became clear that the present invention can estimate a noise probability model having a shape close to the logarithmic mel spectrum histogram of noise as compared with the method disclosed in Non-Patent Document 1. In particular, the results of 804 (L = 3) and 805 (L = 4) show almost the same shape as the histogram of the log mel spectrum of noise.

  FIG. 12 shows the waveform of the audio signal, the waveform of the input acoustic signal (airport lobby noise, S / N ratio 0 dB), and the waveform of the noise suppression signal according to the present invention (L = 1, 2, 3, 4). It can be seen that noise is effectively suppressed by the invention.

  FIG. 13 shows the result of noise suppression. The evaluation result of the speech recognition by the method disclosed in Non-Patent Document 1 and the present invention (L = 1, 2, 3, 4) is shown. From the results shown in FIG. 13, it has been clarified that the present invention can obtain higher performance than the prior art.

<Other embodiments>
In the above embodiments, rectangular window besides Hamming window in the window function w _n at frame cutout processing s201 acoustic feature extraction unit 104, a Hanning window may be used a window function, such as Blackman windows.

  In the above embodiment, other probabilistic models such as HMM (Hidden Markov model) may be used instead of the silent GMM and the clean voice GMM as the voice model stored in advance in the storage unit. Further, as a speech model stored in the storage unit, not only two GMMs, the silent GMM and the clean speech 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 above embodiment, instead of the noise GMM, another probability model such as an HMM may be used as the noise model. At this time, if each state of the HMM is expressed by a mixed normal distribution (GMM) or the like, a noise signal having a multimodal distribution can be modeled as in the first embodiment.

  In the above embodiment, in the noise suppression filter estimation processing s603 of the noise suppression filter estimation unit 501, the estimation result having the maximum weight may be used as it is, instead of the weighted average. In this case, it is desirable to have a sufficiently large weight compared to the weights of other estimation results.

  In the above embodiment, signals are directly input / output between each unit and each means. However, a configuration may be adopted in which signals are stored in a storage unit (not shown) and signals are transferred via the storage unit.

  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.

<Program and recording medium>
The noise suppression device described above can also be functioned by a computer. In this case, each process of a program for causing a computer to function as a target device (a device having the functional configuration shown in the drawings in various embodiments) or a processing procedure (shown in each embodiment) is processed by the computer. A program to be executed by the computer may be downloaded from a recording medium such as a CD-ROM, a magnetic disk, or a semiconductor storage device or via a communication line into the computer, and the program may be 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,
An acoustic feature extraction unit for extracting an acoustic feature of the acoustic signal;
A storage unit for storing a probability model (hereinafter referred to as “speech model”) of a speech signal that does not include noise;
A noise model estimator that estimates a noise signal using the acoustic features of the acoustic signal and the speech model, and performs unsupervised learning of a noise signal probability model (hereinafter referred to as “noise model”) using the estimated noise signal as learning data. When,
A noise suppression unit that suppresses the noise signal of the acoustic signal using a speech model and a noise model;
Including a noise suppression device. The noise suppression device according to claim 1,
The noise signal is defined as a signal based on non-stationary noise that follows a multimodal distribution, and the noise model estimation unit learns the noise model unsupervised.
Noise suppression device. The noise suppression device according to claim 1 or 2,
The noise model estimator is
Probability model generation means for generating a probability model of the acoustic signal (hereinafter referred to as “acoustic model”) using the noise model and the speech model;
Noise signal estimating means for estimating the noise signal using the acoustic signal;
Noise model parameter estimation means for estimating the noise model parameters using the noise signal as learning data,
Using the acoustic signal, the processing of the probability model generation means, the noise signal estimation means, and the noise model parameter estimation means is repeated until a convergence condition is satisfied by an expected value maximization method so that the likelihood of the acoustic model is maximized,
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 extraction step for extracting an acoustic feature of the acoustic signal;
A noise signal is estimated using the acoustic features of the acoustic signal and a noise signal probability model (hereinafter referred to as “voice model”), and the estimated noise signal is used as learning data to obtain a noise signal probability model (hereinafter “ A noise model estimation step for unsupervised learning of a noise model),
A noise suppression step of suppressing a noise signal of the acoustic signal using a noise model;
Including a noise suppression method. The noise suppression method according to claim 4,
The noise signal is defined as a signal based on non-stationary noise following a multimodal distribution, and the noise model is unsupervised learning in the noise model estimation step.
Noise suppression method. The noise suppression method according to claim 4 or 5, wherein
The noise model estimation step includes:
A probability model generation sub-step of generating a probability model of the acoustic signal (hereinafter referred to as “acoustic model”) using the noise model and the speech model;
A noise signal estimation substep for estimating the noise signal using the acoustic signal;
Noise model parameter estimation sub-step for estimating the noise model parameters using the noise signal as training data,
Using the acoustic signal, the probability model generation sub-step, the noise signal estimation sub-step, and the noise model parameter estimation sub-step are repeated until a convergence condition is satisfied by an expected value maximization method so that the likelihood of the acoustic model is maximized. ,
Noise suppression method.   A program for causing a computer to function as the noise suppression device according to claim 1.

Images