JP2008298844A - Noise suppressing device, computer program, and speech recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To suppress noise in a short period of time by using a limited calculation resource in nonsteady noise environment. <P>SOLUTION: A noise suppressing device includes: a noise probability distribution estimation section 200 for generating an estimation parameter of noise distribution for each frame, from a feature amount extracted from the frame of a predetermined period of an observation signal, by using a particle filter; an observation signal distribution estimation section 202 for adapting a Gaussian Mixture Model (GMM) for clean voice estimation according to the estimated noise probability distribution; a clean voice estimation section 204 for calculating an estimated feature amount of object voice for each frame by a Minimum Mean Square Error (MMSE) estimation method; and a calculation control section 212 for controlling an interval of adaptation so that adaptation of the GMM is performed to a plurality of frames at a time. The estimated feature amount of the object voice is calculated by using the Gaussian Mixture Model of a specified frame, when the adaptation is performed to the frame, while by using the Gaussian Mixture Model which is adapted to the previous frame, when the adaptation is not performed. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a speech recognition technique in a real environment where noise is generated, and more particularly to a noise suppression device for improving a speech recognition rate in an environment where non-stationary noise is generated, and speech recognition using the same. About the system.

  Speech recognition technology has been studied as a technology for realizing a human machine interface that is easy and natural for humans. In recent years, speech recognition at a high recognition rate has been realized by a large-scale speech / text database and statistical stochastic speech recognition techniques. Nowadays, the development of applied technology for realizing speech recognition at high speed and with a high recognition rate in an actual environment where a human and a machine are in contact with each other is underway.

  One of the major differences between the actual environment and the laboratory environment is the presence of noise. Noise is generated constantly and irregularly at a volume that cannot be ignored, and fluctuates over time. Noise is a hindrance when performing speech recognition. Improving the speech recognition rate in a real environment where noise is generated is a problem that should be solved as soon as possible in developing an application technology for speech recognition.

  One of the technologies for improving the speech recognition rate in an environment where noise is generated is to estimate and suppress the noise at the pre-processing stage of speech recognition for noise that has a stationary property over time. There is technology.

  Non-Patent Document 1 described later discloses a spectral subtraction method which is a general method for suppressing stationary noise. In this method, it is assumed that the noise amplitude spectrum observed in the section before the utterance is the same as the noise amplitude spectrum in the section during the utterance. Based on this assumption, the noise is suppressed by subtracting the amplitude spectrum of the noise observed immediately before the utterance from the amplitude spectrum of the speech signal observed during the utterance.

  Non-Patent Document 2 described later discloses a noise suppression method in distributed speech recognition. In this method, noise suppression based on the Wiener filter theory is performed using the noise amplitude spectrum observed immediately before the utterance.

  There is also a technique for sequentially estimating and suppressing noise in the preprocessing stage of speech recognition. Non-Patent Document 3 described later discloses a method of sequentially obtaining a maximum likelihood estimation value of noise by applying a sequential EM (Expectation Maximization) algorithm. In the method of sequentially estimating noise using the sequential EM algorithm, noise can be estimated and suppressed with high accuracy while coping with temporal fluctuation of noise.

  Non-patent literature 4 and non-patent literature 5, which will be described later, generally use a method of sequentially obtaining an estimated value of noise using a Kalman filter. In this method, noise is sequentially estimated and suppressed by alternately performing first-term prediction and filtering.

  Further, as a technique for improving the speech recognition rate in a noisy environment, there is a technique for performing adaptive speech recognition using a stochastic model considering noise. For example, in Patent Document 1 described later, noise parameters are estimated using a sequential estimation method called a particle filter, and temporal growth of a hidden state constituting an HMM (Hidden Markov Model) is performed. A speech recognition system that performs speech recognition based on the HMM is disclosed.

S. F. Bol: “Suppression of acoustic noise in speech using spectral subtraction”, IEEE Trans. ASSP, Vol. 27, no. 2, 113-120, 1979 (S.F. Boll: “Suppression of Acoustic Noise in Speech Using Spectral Subtraction,” IEEE Trans. ASSP, Vol. 27, No. 2, pp. 113-120, 1979) European Telecommunications Standards Institute (ETSI) Recommendation ES 202 050 V1.1.3 “Aspects of Speech Processing, Transmission, and Quality (STQ), Distributed Speech Recognition: Advanced Front End Feature Extraction Algorithm; Compression Algorithm ", November 2003 (ETSI ES 202 050 V1.1.3," Speech Processing, Transmission and Quality Aspects (STQ), Distributed Speech Recognition: Advanced Front-end Feature Extraction Algorithm; Compression Algorithms, "Nov. 2003.) M.M. Affifi, O. Shioan: “Sequential estimation with optimal forgetting for robust speech recognition”, IEEE Trans. SAP, Vol. 12, no. 1, 19-26, 2004 (M. Afify, O. Siohan: “Sequential Estimation with Optimal Forgetting for Robust Speech Recognition,” IEEE Trans. SAP, Vol. 12, No. 1, pp. 19-26, 2004. ) Takashi Arimoto: “Kalman Filter”, industrial books Supervised by Michio Nakano, Kiyoshi Nishiyama: “Kalman filter solved on a personal computer”, Maruzen A. M.M. Paynad et al., “Mitigation of channel errors by MMSE for distributed speech recognition”, Euro Speech 2001 Scandinavia (7th European Conference on Speech Communication and Technology), pp. 2707-2710, 2001 (Peinado AM, Sanchez V, Segura JC, Perez-Cordoba JL, "MMSE-Based Channel Error Mitigation for Distributed Speech Recognition," Eurospeech 2001-Scandinavia (7th European Conference on Speech Communication and Technology), pp .2707-2710, 2001) JP 2007-41499 A

  The techniques described in Non-Patent Document 1 and Non-Patent Document 2 are both techniques for estimating and suppressing noise on the assumption that the noise is stationary. However, most of the noise in the real environment is non-stationary. That is, the acoustic characteristics of noise vary with time. For this reason, the techniques described in Non-Patent Document 1 and Non-Patent Document 2 cannot cope with temporal fluctuations in noise and cannot suppress noise with high accuracy.

  In the technique described in Non-Patent Document 3, a sequential EM algorithm is used. When noise is estimated by the sequential EM algorithm, it is necessary to perform iterative calculation for each frame of the observed speech signal until the parameters in the frame converge to the local optimum value of the likelihood function. Therefore, an enormous amount of calculation is required every time the noise fluctuates, and the calculation takes time. Therefore, it is difficult to estimate and suppress noise in real time by this method.

  In the techniques described in Non-Patent Document 4 and Non-Patent Document 5, noise is estimated using a Kalman filter. This estimation method is a method of alternately performing first-term prediction and filtering, and does not require iterative calculation like a sequential EM algorithm. However, the method using the Kalman filter estimates the probability distribution assuming that the posterior probability distribution of noise is a single normal distribution. When the true posterior probability distribution is a mixed distribution, it is approximated by a single normal distribution. Therefore, the accuracy is deteriorated.

  In the speech recognition system described in Patent Document 1, since speech recognition is performed using a model that takes noise into account, an effect of increasing speech recognition accuracy is obtained. However, since the particle filter is used, the calculation amount is large, and it is difficult to operate the system at high speed with an apparatus having limited calculation resources.

  SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a noise suppression device that can improve a speech recognition rate in an environment where non-stationary noise occurs and can suppress noise in a short time using limited calculation resources. It is to be.

  Another object of the present invention is to provide a noise suppression device that can improve a speech recognition rate in an environment where non-stationary noise occurs and can suppress noise in a short time while reducing the amount of calculation. is there.

  A noise suppression apparatus according to a first aspect of the present invention is a noise suppression apparatus for suppressing a noise component in an observation signal obtained by observation of a target speech in an environment where noise is generated. A clean speech estimation comprising a plurality of element distributions prepared in advance using a particle filter having a plurality of particles, each receiving a feature amount extracted from a frame of a predetermined time length framed at a predetermined period. Based on the acoustic model for the noise, the noise estimation means for sequentially generating the estimation parameter of the probability distribution representing the noise for each frame, and the noise model according to the noise probability distribution whose parameter is estimated by the noise estimation means Using the adaptation means for adapting to the noise, the acoustic model adapted to noise, and the feature quantity of the observed signal. The target speech estimation means for calculating the estimated feature amount of the target speech by the MMSE (Minimum Mean Square Error) estimation method and the adaptation means for adaptation of the acoustic model for each frame And a control unit for controlling the interval of adaptation by the adaptation unit, wherein the target speech estimation unit performs adaptation by the adaptation unit for a certain frame. The estimated feature quantity of the target speech is calculated for the frame using the adapted Gaussian mixture model, and when the adaptation means is not applied to a frame, For the frame, the estimated feature quantity of the target speech is calculated using the acoustic model adapted by the adaptation means for the previous frame. That.

  Details of the MMSE method are disclosed in Non-Patent Document 6.

  In this specification, the voice (observation voice) given to the noise suppression device is considered as a sound in which the target voice without noise and noise are superimposed. The target voice when considered in this way is called “clean voice”.

  When the adaptation unit adapts the acoustic model for clean speech estimation to noise, the control unit controls the adaptation interval so that the adaptation is performed for each of a plurality of frames. When adaptation is performed for a certain frame, the target speech estimation means estimates the target speech by the MMSE estimation method using the acoustic model. When the adaptation is not performed for a certain frame, the target speech estimation means estimates the target speech by the MMSE estimation method using the acoustic model adapted to the previous frame. A process with a large amount of calculation such as adaptation of the acoustic model need not be performed on each frame, and the amount of calculation is reduced. Therefore, the target speech can be estimated at high speed.

  The acoustic model may be a Gaussian mixture distribution including a plurality of element distributions.

  If the Gaussian mixture distribution is used, even if the feature amount of clean speech has a complicated distribution, it becomes easy to model it statistically.

  Any two of the plurality of element distributions may have different variances.

  In general, since the element distribution is statistically obtained by learning on the speech sample, the variance is often different from each other.

  The plurality of element distributions may have different averages and equal variances.

  In this way, it is virtually unnecessary to consider the distribution in the calculation in the noise estimation means, and the amount of calculation is further reduced. As a result, processing can be made faster. As a result, it is possible to provide a noise suppression device capable of improving a speech recognition rate in an environment where non-stationary noise occurs and capable of suppressing noise in a short time using limited calculation resources. Can do.

  Preferably, the noise suppression device further includes storage means for storing the adapted acoustic model in response to the adaptation of the acoustic model for a frame by the adaptation means, and the target speech When the acoustic model for clean speech estimation is adapted for a certain frame, the estimating means calculates the estimated feature amount of the target speech by using the adapted acoustic model and the MMSE estimation method. When the acoustic model for the clean speech estimation is not adapted for, the estimated feature quantity of the target speech is calculated by the MMSE method using the adapted acoustic model stored in the storage means. To do.

  When the acoustic model is adapted for a certain frame, the acoustic model is stored in the storage means. When the acoustic model is not adapted for a certain frame, the estimated feature quantity of the target speech is calculated by the MMSE estimation method using the acoustic model stored in the storage means. When the adaptation is not performed for a certain frame, the estimated feature amount of the target speech is calculated using the acoustic model adapted for the previous frame. Since this process is repeated in a timely manner, even if an acoustic model adapted for the previous frame is used in this way, if the interval between adaptations is sufficiently short, the target speech There is almost no performance impact on the estimation of. On the other hand, the calculation amount for adaptation can be greatly reduced. As a result, the target speech can be estimated at high speed while maintaining the performance.

  More preferably, the control means includes an interval storage means for storing information for determining an interval between frames for performing processing for adapting an acoustic model for clean speech estimation to noise in advance, and an adaptation means immediately before Frame number storage means for storing the number of frames processed after adaptation, and for adding 1 to the stored contents of the frame number storage means every time a frame to be processed is given to the noise suppression device An adder, a determination unit for determining whether the stored contents of the frame number storage unit and the stored content of the interval storage unit are equal, and adaptation by the adaptation unit according to the determination result by the determination unit And a means for performing adaptation processing for disabling adaptation to the frame by the adaptation means, and a decision means In response to the stored contents of the memory contents and spacing storage means beam number storage means is determined to be equal, and means for clearing the frame number memory means to zero.

  When the computer program according to the second aspect of the present invention is executed by a computer, it causes the computer to operate as any of the noise suppression devices described above. Therefore, the same effect as the above-described noise suppression device can be obtained.

  The speech recognition system according to the third aspect of the present invention receives any of the above-described noise suppression device and the estimated feature amount of the target speech calculated by the noise suppression device, and receives the above-described acoustic model and recognition target language. And speech recognition means for performing speech recognition on the target speech using a predetermined language model.

[First Embodiment]
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 parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed description thereof will not be repeated. The symbol “^” or the like used in the text of the following description should be described immediately above the character immediately after it, but it is described immediately before the character due to restrictions on text notation. In the formula, these symbols are written in their original positions. Also, in the text of the following description, in the case of a vector or a matrix, in many cases, for example, a normal text immediately preceded by “vector”, “matrix”, etc. like “vector X _t ”, “matrix Σ _W ”, etc. In the formula, all are written in bold.

〔Constitution〕
<Configuration of the entire speech recognition system>
FIG. 1 shows the overall configuration of the speech recognition system 100 according to the present embodiment. With reference to FIG. 1, the speech recognition system 100 includes a preprocessing unit 104 for extracting a feature vector 126 representing a feature of speech used for speech recognition from a sound 122 generated by a sound source 102, and a preprocessing unit 104. And a pre-processing acoustic model unit 106 for preparing a probabilistic model (acoustic model) representing the relationship between speech features and phonemes, and a probabilistic model (language for expressing word connection probabilities in a language to be recognized) A model for preparing a model), and a search for searching for a word or the like corresponding to the feature amount output from the preprocessing unit 104 using the language model of the language model unit 108 and a predetermined acoustic model. And a recognition acoustic model unit 109 for preparing an acoustic model connected to the search unit 110 and used for the search by the search unit 110.

  The speech recognition system 100 further includes a constraint condition parameter 138 that is used for the extraction of the feature vector 126 by the preprocessing unit 104 and includes a coefficient for determining a constraint condition in a state space model described later.

  The sound source 102 includes a speaker 116 that utters a speech (target speech) 120 to be recognized, and a noise source 118 that generates noise 121 around the speaker 116. The sound 122 generated by the sound source 102 and recorded by the preprocessing unit 104 is a sound in which the target voice 120 without noise and noise 121 generated by the speech of the speaker 116 are superimposed. In this specification, as described above, the target voice 120 that does not include noise is referred to as “clean voice”. On the other hand, the sound 122 that reaches the preprocessing unit 104 and is recorded by the preprocessing unit 104, that is, the sound 122 in which the clean sound 120 and the noise 121 are superimposed is referred to as “noise superimposed sound”.

  The pre-processing unit 104 records the noise-superimposed speech 122 and performs predetermined signal processing on the observation signal obtained as a result, thereby obtaining a predetermined feature vector related to the observation signal (hereinafter referred to as “observation signal”). And a noise component included in the feature value 124 of the observation signal extracted by the measurement unit 112 is extracted from the measurement unit 112 for extracting 124 and the preprocessing acoustic model unit 106. And a noise suppression unit 114 for suppression using the acoustic model prepared by the above and the constraint condition parameter 138.

  The measurement unit 112 analyzes the logarithmic mel filter bank of the observation signal for each frame having a frame interval of 10 milliseconds and a time length of several tens of milliseconds, and uses the obtained logarithmic mel spectrum as an element as a feature 124 Output as.

  The noise suppression unit 114 uses the acoustic model prepared by the preprocessing acoustic model unit 106 and the constraint condition parameter 138 to generate the feature vector of the clean speech 120 for each frame based on the feature 124 of the observation signal. Has the ability to estimate sequentially. The feature quantity vector obtained by this successive estimation is output to the search unit 110 as a voice feature quantity vector 126 used for speech recognition. At this time, first, the feature vector of the noise 121 is estimated, and the feature vector of the clean speech 120 is estimated based on the result. In this specification, the voice represented by the feature vector 126 is referred to as “estimated clean voice”. Also, the feature quantity vector 126 is referred to as “estimated clean speech feature quantity”.

  The search unit 110 uses the feature quantity 126 of the estimated clean speech, and based on the acoustic model prepared by the recognition acoustic model unit 109 and the language model prepared by the language model unit 108, A word or the like is searched, and the result is output as a recognition output 128.

<Acoustic model for pretreatment>
Hereinafter, an acoustic model prepared by the preprocessing acoustic model unit 106 will be described. The pre-processing acoustic model unit 106 shown in FIG. 1 prepares and holds a Gaussian Mixture Model (GMM) 130 as an acoustic model for the clean speech 120. The preprocessing acoustic model unit 106 performs learning for the GMM 130 using the learning data storage unit 132 for storing learning data related to the clean speech 120 prepared in advance and the learning data in the learning data storage unit 132. A model learning unit 134 and a GMM storage unit 136 for storing the GMM 130 obtained by learning by the model learning unit 134 are included.

FIG. 2 schematically shows the concept of the GMM 130. Referring to FIG. 2, the GMM 130 is a probability model in which a time series signal is modeled by one stationary signal source (state). In GMM 130, a vector that may be output as a feature vector of clean speech 120 and a probability that the vector is output (hereinafter simply referred to as “output probability”) are defined. The output probability is expressed by a mixed normal distribution 140. The mixed normal distribution 140 in the GMM 130 includes a plurality of element distributions 148A, 148B,. These element distributions 148A, 148B, ..., 148K are all single normal distributions. For example, certain elements distribution 150 contained in the mixed normal distribution 140 and k _t. The element distribution k _t is a single normal distribution, and considering multidimensionality, the distribution average vector μ _{S, kt} (hereinafter simply referred to as “average”) and the covariance matrix Σ _{S, kt} (hereinafter simply referred to as “dispersion”). "). All of these are statistically learned (calculated) based on various voice samples in advance. A vector of parameters output with a probability according to the element distribution k _t 150 is set as a vector S _{kt, t} . In the following description, the parameter vector S _{kt, t} output from the GMM 130 is referred to as “output parameter (of the GMM 130)”.

<State space model>
The state space model will be described below. A state space model is a dynamic model consisting of an observation equation that represents the generation process of an observed signal and a state equation that represents a process that changes the processing target (hereinafter, this process is referred to as a “state transition process”). It is. FIG. 3 schematically shows the state space model 160.

A feature quantity 124 (see FIG. 1) of the observation signal in the frame at time t (hereinafter simply referred to as “tth frame”) is represented by X _t . The feature amount X _t of the observation signal is a vector having a log mel spectrum obtained from the noise superimposed speech 122 as an element as described above. The feature quantity X _t of the observation signal has a logarithmic Mel spectra of clean speech 120 and noise 121 and the sound of superimposed elements. Here, the feature vector S _t of the clean speech vector with a logarithmic Mel spectra of clean speech 120 in the t frame element. A vector having the log mel spectrum of the noise 121 as an element is defined as a noise feature vector N _t . The dimensions of the vectors X _t , S _t and N _t are the same. Note that the processing described below is performed for each element of the vector and matrix, but in the following description, each element is not particularly distinguished for the sake of simplicity.

First, an observation signal generation process in the state space model 160 will be described. Feature quantity X _t of the observation signals is a known vector obtained by the measurement. On the other hand, the clean speech feature vector _St and the noise feature vector _Nt are unknown vectors that cannot be obtained by measurement.

Here, it is assumed that the output process of the clean speech 120 can be modeled by GMM. That is, it is assumed feature vector S _t of the clean speech in the t frame, the output parameter vector S _kt outputted in accordance with the GMM130 element distribution k _t 0.99 (see FIG. _2), as represented by _t . However, there is an error between the clean speech feature vector _St and the output parameter vector _{Skt, t} . This error is also a vector. This error and error vector V _t. As shown in the following equation, the error vector V _t has a value according to a probability distribution represented by a single normal distribution having an average of 0 and a variance of Σ _{S, kt} as an element.

ただし、この式においてΣ_S,ktはＧＭＭ１３０内のある要素分布ｋ_t１５０より得られるパラメータの共分散行列を表し、記号「〜」は左辺の値が右辺に示される確率分布にしたがうことを示す。すなわち、左辺の値が右辺に示す確率分布にしたがったサンプリングにより推定できることを示す。また、この式において、「Ｎ（μ，Σ）」は、平均がμで分散がΣの単一正規分布を表す。

In this equation, Σ _{S, kt} represents a covariance matrix of parameters obtained from a certain element distribution k _t 150 in the GMM 130, and the symbol “˜” indicates that the value on the left side follows the probability distribution shown on the right side. . That is, the value on the left side can be estimated by sampling according to the probability distribution shown on the right side. In this equation, “N (μ, Σ)” represents a single normal distribution with an average of μ and a variance of Σ.

Based on the above assumption, the generation process of the observed signal feature quantity X _t 124 is performed using the noise feature quantity vector N _t , the output parameter vector S _{kt, t} from the GMM 130, and the error vector V _t as follows: It shall be expressed by the observation equation shown in (1).

なお、式（１）でＩは単位ベクトルを表す。また、ベクトルの対数、ベクトルの指数演算はそれぞれ、ベクトルの各要素について対数をとり、又は指数計算し、その結果を成分とするベクトルを表すものとする。

In Equation (1), I represents a unit vector. In addition, the logarithm of the vector and the exponent operation of the vector respectively represent a vector having a logarithm or exponent calculation for each element of the vector and using the result as a component.

Next, a state transition process to be processed in the state space model 160 will be described. In the state space model 160, the noise of the feature vector N _t is the target of processing. Here, it is assumed that the noise feature vector N _t changes according to the random walk process. That is, it is assumed feature vector N _t-1 noise in the t-1 frame is between the noise feature vector N _t in the t frame, as random changes occur. A vector representing this random change is a random Gaussian noise vector W _t . Random Gaussian noise vector W _t is assumed average is random Gaussian noise with a value according to the probability distribution that is expressed to the element by a single normal distribution variance sigma _w 0.

ただし、この式においてΣ_Wは、ランダムガウス雑音ベクトルＷ_tの共分散行列を表す。

In this equation, Σ _W represents the covariance matrix of the random Gaussian noise vector W _t .

When a state equation expressing the state transition process of the noise feature vector N _t is defined based on the above assumption, the state equation is expressed as the following equation (2).

しかし、ランダムウォーク過程に基づく上記の仮定では、雑音の特徴量ベクトルＮ_tの変化をランダムガウス雑音ベクトルＷ_tで規定している。そのため、式（２）に示す状態方程式では、雑音の特徴量ベクトルＮ_tの時間変化を正確に表現することはできない。そこで、本実施の形態では、図１に示す拘束条件パラメータ１３８を用いて、雑音の特徴量ベクトルＮ_tの変化に対し、拘束条件を設ける。拘束条件及びそのための拘束条件パラメータ１３８の詳細については、後述する。

However, the above assumption based on the random walk process, defines a random Gaussian noise vector W _t changes in the noise feature vector N _t. For this reason, the state equation shown in Expression (2) cannot accurately represent the temporal change of the noise feature vector N _t . Therefore, in the present embodiment, a constraint condition is provided for a change in the noise feature quantity vector N _t using the constraint condition parameter 138 shown in FIG. Details of the constraint condition and the constraint condition parameter 138 for the constraint condition will be described later.

<Configuration of Noise Suppression Unit 114>
FIG. 4 is a block diagram showing the configuration of the noise suppression unit 114 (see FIG. 1). Referring to FIG. 4, the noise suppression unit 114, feature quantity of the observation signal X _t 124, GMM130, and using a constraint enforcement parameter 138, the probability distribution representing the probability of the output of the noise feature vector N _t ( (Hereinafter referred to as “noise probability distribution”) is sequentially estimated for each frame, and a parameter (hereinafter referred to as “estimation parameter of noise probability distribution”) 205 representing the noise probability distribution is generated. Noise probability distribution estimator 200, noise probability distribution estimation parameter 205 estimated by noise probability distribution estimator 200, and GMM 130, a probability distribution that represents the probability of output of observed signal feature quantity X _t 124 ( Hereinafter, the parameter 208 representing the observation signal probability distribution (hereinafter referred to as “observation signal distribution parameter”) is estimated. And a monitoring signal distribution estimation unit 202 for generating the called.) And.

  The noise suppression unit 114 further includes an interval storage unit 206 for preliminarily storing a frame interval (“L”) at which the observation signal distribution is estimated, and the noise probability distribution estimation unit 200 to estimate the noise probability distribution estimation parameter 205. , The calculation control unit 212 for controlling the observation signal distribution estimation unit 202 so as to estimate the observation signal distribution every L-th frame, and the calculation control unit 212 controls the observation. And a number storage unit 218 used for counting the number of frames for which the signal distribution has not been estimated.

The calculation control unit 212 has a function of providing the observation signal distribution estimation unit 202 with an estimation instruction signal 209 that indicates the timing of a frame for which the observation signal distribution is to be estimated, and the observation signal distribution estimation unit 202 estimates a certain frame. The function of outputting the observation signal distribution storage control signal 207 indicating the timing at which the output observation signal distribution should be stored, and the first value when it is the timing of the frame for which the observation signal distribution is to be estimated, and otherwise And a function of outputting a selection instruction signal 211 having a second value. That is, the calculation control unit 212 adds 1 to the value I _L stored in the frequency storage unit 218 each time to start processing of one frame is compared to the value L stored in the distance storage unit 206 the value Has function. The calculation control unit 212 outputs the observation signal distribution storage control signal 207 and the estimation instruction signal 209 when they become equal. At this time, the calculation control unit 212 also sets the value of the selection instruction signal 211 as the first value. In other cases, the calculation control unit 212 sets the value of the selection instruction signal 211 as the second value. The calculation control unit 212 resets the value of the number storage unit 218 to zero when the value of the number storage unit 218 becomes equal to L.

  The noise suppression unit 114 further includes a distribution storage unit 214 for storing and outputting the observation signal distribution estimated by the observation signal distribution estimation unit 202 in response to the observation signal distribution storage control signal 207 from the calculation control unit 212. , Having two inputs respectively connected to the output of the distribution storage unit 214 and the output of the observation signal distribution estimation unit 202, and when the selection instruction signal 211 is the first value, the parameter 208 of the observation signal distribution Sometimes the last estimated observation signal distribution parameter 213 is selected and output as the observed signal distribution parameter 217, the observed signal feature 124, and the observed signal distribution output by the selecting unit 216. Clean speech estimation for generating the estimated clean speech feature 126 based on the parameters 217 and the GMM 130 And a 204.

The noise probability distribution estimation unit 200 has a function of sequentially estimating the noise probability distribution for each frame and outputting an estimation parameter 205 of the noise probability distribution. Here, a series of vectors composed of observed signal feature values X ₀ ,..., X _t is a sequence X _{0: t} = {X ₀ ,..., X _t }, and noise feature vector N _{0 ,.} A vector sequence consisting of the sequence N _{0: t} = {N ₀ ,..., N _t }. The posterior probability distribution p (N _{0: t} | X _{0: t} ) of the sequence N _{0: t} when the observation signal vector sequence X _{0: t} is given is _expressed by the following equation (1) using a first-order Markov chain _: It is expressed as 3).

Therefore, the problem of successively estimating the probability distribution of the noise feature vector N _t is to maximize the posterior probability p (N _{0: t} | X _{0: t} ) when the observation signal vector sequence X _{0: t} is given. This results in the problem of estimating the sequence N _{0: t} . The noise probability distribution estimation unit 200 performs this estimation based on the observed signal feature amount X _t 124, the GMM 130, the state space model 160, and the constraint condition parameter 138 regarding the noise state transition. At that time, the noise probability distribution estimation unit 200 uses a technique called a particle filter. This estimation method generates many localized state spaces (particles) in a state space represented by a certain state space model, estimates the probability distribution of parameters for each particle, and uses each particle to This is a technique for approximately expressing the probability distribution of parameters in space.

  The calculation control unit 212 uses the frame interval (L) stored in the interval storage unit 206 and the number-of-times storage unit 218 to estimate the parameters of the observation signal distribution by the observation signal distribution estimation unit 202 for each frame. And has a function of controlling the observation signal distribution estimation unit 202 so as to be performed every L frames. The calculation control unit 212 further provides a function of giving the observation signal distribution storage control signal 207 to the distribution storage unit 214 so as to store the value when the observation signal distribution parameter 208 is estimated by the observation signal distribution estimation unit 202, A selection instruction signal 211 that takes a first value for a frame in which the observation signal distribution parameter 208 is estimated by the observation signal distribution estimation unit 202 and a second value for the other frames. Is output to the selection unit 216.

  The observation signal distribution estimation unit 202 has a function of calculating an average vector and a covariance matrix of the observation signal distribution in each particle as the observation signal distribution parameter 208. For example, an HMM synthesis method called a VTS (Vector Taylor Series) method is used to calculate the parameter 208 of the observation signal distribution.

  When the observation signal distribution storage control signal 207 is given from the calculation control unit 212, the distribution storage unit 214 stores the observation signal distribution parameter 208 output from the observation signal distribution estimation unit 202 at that time. The output of the distribution storage unit 214 is given to one input of the selection unit 216 as the parameter 213 of the observed signal distribution estimated last.

  The selection unit 216 receives the observation signal distribution parameter 208 from the observation signal distribution estimation unit 202 when the selection instruction signal 211 provided from the calculation control unit 212 is the first value, and from the distribution storage unit 214 when the selection instruction signal 211 is the second value. The parameters 213 of the observed signal distribution estimated at the end of each are selected and given to the clean speech estimation unit 204.

  The clean speech estimation unit 204 has a function of estimating clean speech parameters for each particle for each frame and calculating a feature amount 126 of the estimated clean speech by a weighted sum of these estimated parameters. For example, the MMSE estimation method is used to calculate the feature amount 126 of the estimated clean speech. The clean speech estimation unit 204 further has a function of issuing a request 210 regarding the transition to the next frame to the noise probability distribution estimation unit 200.

<Particle filter>
Hereinafter, the particle filter will be described. In this method, initial parameters in a large number of particles are determined by random sampling or sampling from a probability distribution representing the initial state of the parameters. Then, the following processing is performed for each frame. That is, when a parameter is determined for each particle corresponding to a certain frame, first, the parameter of each particle is updated to one corresponding to a frame subsequent to the frame. Subsequently, a weight is assigned to each particle according to the likelihood of update. Subsequently, the parameter of each particle corresponding to the subsequent frame is resampled according to the parameter probability distribution in the updated particle. Subsequently, the parameter of each particle corresponding to the subsequent frame is determined based on the resampled parameter. By performing the above processing for each frame, parameters for each particle are sequentially determined.

In the particle filter, each parameter in the state space model 160 is approximately expressed by a weighted sum of parameters in the particle. Here, the number of particles is J, and the noise feature vector of the j (1 ≦ j ≦ J) -th particle in the t-th frame is a vector N _t ^(j) . Further, let w _t ^(j) be the weight for the j-th particle in the t-th frame. The posterior probability distribution p (N _{0: t} | X _{0: t} ) shown in the equation (3) is approximately expressed by the Monte Carlo sampling shown in the following equation (4).

なお、この式においてδ（）は、Dirac-delta関数を表す。

In this equation, δ () represents the Dirac-delta function.

If the probability distribution for outputting the noise feature vector series N _{0: t} ^(j) in the j-th particle is q (N _{0: t} ^(j) | X _{0: t} ), the weight w _t ^{(j )} Is given by the following equation (5).

確率分布ｑ（Ｎ_0:t ^(j)｜Ｘ_0:t）は、次の式（６）に示す連鎖モデルで表現されるものとする。

The probability distribution q (N _{0: t} ^(j) | X _{0: t} ) is assumed to be expressed by a chain model shown in the following equation (6).

また、上記の式（３）の事後確率分布ｐ（Ｎ_0:t｜Ｘ_0:t）は、ベイズ則により次の式（７）のように表現できる。

Further, the posterior probability distribution p (N _{0: t} | X _{0: t} ) of the above equation (3) can be expressed as the following equation (7) by Bayes rule.

したがって、式（５）、式（６）、及び式（７）より、パーティクルに対する重みｗ_t ^(j)は、式（８）によって与えられることになる。

Therefore, the weight w _t ^(j) for the particles is given by the equation (8) from the equations (5), (6), and (7).

ここで、ｐ（Ｎ_t ^(j)｜Ｎ_t-1 ^(j)）＝ｑ（Ｎ_t ^(j)｜Ｎ_0:t-1 ^(j)，Ｘ_0:t）と仮定すると、式（８）より、式（９）が得られる。

Assuming that p (N _t ^(j) | N _t-1 ^(j) ) = q (N _t ^(j) | N _{0: t-1} ^(j) , X _{0: t} ), the equation (8 ), Equation (9) is obtained.

式（９）のｐ（Ｘ_t｜Ｎ_t ^(j)）は、次の式（１０）に示す確率密度関数によりモデル化される。

P (X _t | N _t ^(j) ) in the equation (9) is modeled by a probability density function shown in the following equation (10).

The noise probability distribution estimation unit 200 uses, as the noise probability distribution estimation parameter 205, the probability density function in the equation (4) for the feature vector N _t ^(j) of the noise in the particle for each particle j (1 ≦ j ≦ J). The parameter of p (N _{0: t} ^(j) | X _{0: t} ) and the weight w _t ^(j) for the particle are sequentially calculated based on the state space model 160 shown in FIG. The parameters of the probability density function p (N _{0: t} ^(j) | X _{0: t} ) are the mean vector ^ N _t ^(j) of the noise feature vector N _t ^{(j) in} the particle and the covariance matrix Σ _Nt ^(j) . Hereinafter, the mean vector ^ N _t ^{(j) of} the probability density function p (N _{0: t} | X _{0: t} ⁾ and the covariance matrix Σ _Nt ^(j) are expressed as “the noise parameter in the (j-th) particle”. Call it.

<Restrictions for state transition process>
As described above, in the state equation shown in equation (2), it is impossible to accurately represent the time variation of noise feature vector N _t. Therefore, in the present embodiment, the state equation shown in the following equation (11) is introduced with respect to the change of the noise feature vector N _t ^(j) (1 ≦ j ≦ J) in each particle.

この状態方程式（１１）において第１項と第２項とは、第ｔ＋１フレームにおけるパーティクルの散らばりを抑制するための拘束条件である。以下この拘束条件を第１の拘束条件と呼ぶ。また、状態方程式（１１）において第３項は、ｊ番目のパーティクルにおける雑音の特徴量ベクトルの時間推移に対する拘束条件である。以下、この拘束条件を第２の拘束条件と呼ぶ。

In this state equation (11), the first term and the second term are constraint conditions for suppressing particle scattering in the (t + 1) th frame. Hereinafter, this constraint condition is referred to as a first constraint condition. In the state equation (11), the third term is a constraint condition for the time transition of the noise feature vector in the j-th particle. Hereinafter, this constraint condition is referred to as a second constraint condition.

  In the state equation (11), α is a forgetting factor, and β is a scaling factor for the second constraint condition.

In the first constraint, the vector ^ N _t is a weighted average of noise feature vectors N _t ⁽¹⁾ ,..., N _t ^(J) in each particle of the t-th frame, and the following equation (12) Given by.

すなわち、第１の拘束条件により、各パーティクルにおける雑音の特徴量ベクトルは、加重平均ベクトル＾Ｎ_tに近づくよう補正される。

That is, according to the first constraint condition, the feature vector of noise in each particle is corrected so as to approach the weighted average vector ^ N _t .

In the second constraint condition, the vector μ _Nt ^(j) is an average of noise feature vectors N _{t−T + 1} ^(j) ,..., N _t ^{(j) for} the past T frames in the j-th particle ( Polyak Average), which is given by the following equation (13).

すなわち、第２の拘束条件により、パーティクルにおける雑音の特徴量ベクトルにそれぞれ、そのパーティクルにおけるPolyak Averageベクトルμ_Nt ^(j)がフィードバックされる。本実施の形態では、式（１１）に示す状態方程式の忘却係数α及び第２の拘束条件に対するスケーリング係数βと、式（１３）におけるフレーム数Ｔとが、図１に示す拘束条件パラメータ１３８として与えられる。

That is, the Polyak Average vector μ _Nt ^{(j) of the} particle is fed back to the noise feature vector of the particle by the second constraint condition. In the present embodiment, the forgetting factor α of the state equation shown in Equation (11), the scaling factor β for the second constraint condition, and the frame number T in Equation (13) are used as the constraint parameter 138 shown in FIG. Given.

  The noise probability distribution estimation unit 200 sequentially estimates the noise probability distribution using a particle filter based on the state space model represented by the observation equation (1) and the state equation (11).

<Configuration of Noise Probability Distribution Estimation Unit 200>
FIG. 5 is a block diagram showing the configuration of the noise probability distribution estimation unit 200. Referring to FIG. 5, noise probability distribution estimation section 200 receives request 210 from clean speech estimation section 204, selects a frame to be processed from feature quantity 124 of the observation signal, and performs observation corresponding to the frame. A frame selection unit 220 for providing the signal feature quantity 124 to the output destination corresponding to the frame is included.

  The noise probability distribution estimation unit 200 further receives a feature quantity 124 of the observation signal from the frame selection unit 220, estimates a probability distribution representing noise in an initial state (hereinafter referred to as “noise initial distribution”), and For a large number (J) of particles, a noise initial distribution estimation unit 222 for determining a noise probability distribution estimation parameter 205 in a frame at t = 0 (hereinafter, this frame is referred to as an “initial frame”), a frame A sequential calculation unit 224 for sequentially calculating an estimation parameter 205 of the noise probability distribution in the t (t ≧ 1) -th frame for each particle in response to the feature value 124 of the observation signal from the selection unit 220; .

The frame selection unit 220 sequentially selects frames to be processed every time the request 210 is given. When the initial frame is selected as a processing target, the frame selection unit 220 gives the initial predetermined frame (for example, 10 frames) of the observed signal feature amount X _t 124 to the noise initial distribution estimation unit 222. In addition, when the other frame (t ≧ 1) is selected as the processing target, the frame selection unit 220 gives the feature amount X _t 124 of the observation signal in the frame to the sequential calculation unit 224.

  The initial noise distribution estimation unit 222 estimates the parameters of the initial noise distribution as follows.

That is, the initial noise distribution estimation unit 222 estimates the initial noise distribution by regarding the initial noise distribution as a single normal distribution. An initial value vector of noise is a vector N ₀ , and an initial noise distribution is p (N ₀ ). When the average vector in the initial noise distribution p (N ₀ ) is μ _N and the covariance matrix is a matrix Σ _N , the initial noise distribution p (N ₀ ) is expressed as the following equation (14).

雑音初期分布推定部２２２は、最初の所定フレーム分の区間の観測信号の特徴量Ｘ_t１２４が雑音１２１の成分のみからなるものとみなし、式（１４）に示す雑音初期分布ｐ（Ｎ₀）の平均ベクトルμ_Nと共分散行列Σ_Nとを推定する。例えば、０≦ｔ≦９の１０フレーム分の区間が雑音１２１の成分のみからなる区間に該当する場合、雑音初期分布推定部２２２は、平均ベクトルμ_Nと共分散行列Σ_Nとをそれぞれ、次の式（１５）と式（１６）とによって算出する。ただし、式（１６）においてベクトルの右肩に付した「Ｔ」は転置を表す。

The initial noise distribution estimation unit 222 considers that the feature amount X _t 124 of the observed signal in the first predetermined frame interval is composed only of the noise 121 component, and the initial noise distribution p (N ₀ ) shown in Expression (14). Of the mean vector μ _N and the covariance matrix Σ _N. For example, when a section of 10 frames of 0 ≦ t ≦ 9 corresponds to a section including only the noise 121 component, the noise initial distribution estimation unit 222 applies the average vector μ _N and the covariance matrix Σ _N to the next (15) and (16). However, “T” attached to the right shoulder of the vector in Expression (16) represents transposition.

そして雑音初期分布推定部２２２は、初期フレーム（ｔ＝０）でのｊ番目のパーティクルにおける雑音のパラメータであるベクトルＮ₀ ^(j)と共分散行列Σ_N0 ^(j)とを、それぞれ、式（１７）及び式（１８）のように設定する。

Then, the initial noise distribution estimation unit 222 obtains a vector N ₀ ^(j) and a covariance matrix Σ _N0 ^(j) , which are noise parameters of the j-th particle in the initial frame (t = 0), by the formula ( 17) and the equation (18).

すなわち、雑音初期分布推定部２２２は、ｊ番目のパーティクルにおける雑音の特徴量ベクトルＮ₀ ^(j)を、初期分布ｐ（Ｎ₀）からのサンプリングによって生成し、共分散行列Σ_N0 ^(j)を、初期分布ｐ（Ｎ₀）の共分散行列Σ_Nに設定する。雑音初期分布推定部２２２は、式（１７）と式（１８）とに示す設定をパーティクルｊ（１≦ｊ≦Ｊ）ごとに行なう。

That is, the noise initial distribution estimation unit 222 generates a noise feature vector N ₀ ^(j) in the j-th particle by sampling from the initial distribution p (N ₀ ), and generates a covariance matrix Σ _N0 ^(j) . , Set to the covariance matrix Σ _N of the initial distribution p (N ₀ ). The initial noise distribution estimation unit 222 performs the setting shown in Expression (17) and Expression (18) for each particle j (1 ≦ j ≦ J).

  The sequential calculation unit 224 includes a GMM sampling unit 226 for sampling the output parameter 240 from the GMM 130. The sequential calculation unit 224 further receives the feature quantity 124 of the observation signal and updates the noise parameter of each particle, an update unit 230 for calculating the weight for the updated particle, and a weight calculation unit 232 for calculating the weight for the updated particle. Based on the calculated weight, the re-sampling unit 234 for re-sampling the noise parameter in the particle, and determining the noise parameter in each particle based on each re-sampled particle and each particle in the t-1 frame. And an estimation parameter generation unit 236 for generating an estimation parameter 205 of the noise probability distribution.

For each particle j (1 ≦ j ≦ J), the GMM sampling unit 226 calculates an element distribution k _t ^(j) corresponding to the particle from the mixture distribution 140 in the GMM 130 (see FIG. 2) based on the mixture weight. Sampling. Further, the GMM sampling unit 226 samples the output parameter vector S ^(j) _kt ^(j) _{, t} from the element distribution k _t ^(j), and supplies it to the update unit 230. Here, if the mixing weights of the element distributions 148A,..., 148K in the GMM 130 are P _{S, kt} , the element distribution k _t ^(j) follows a probability distribution with the mixing weights P _{S, kt} as output probabilities. That is, it is obtained from the GMM 130 by sampling shown in the following equation (19).

要素分布ｋ_t ^(j)の平均ベクトルをベクトルμ_kt ^(j)とし、要素分布ｋ_t ^(j)の共分散行列を行列Σ_S,kt ^(j)とすると、ｊ番目のパーティクルにおけるＧＭＭ１３０の出力パラメータベクトルＳ^(j) _kt ^(j) _,tは、要素分布ｋ_t ^(j)から、次の式（２０）に示すサンプリングによって得られる。

If the average vector of the element distribution k _t ^(j) is the vector μ _kt ^(j) and the covariance matrix of the element distribution k _t ^(j) is the matrix Σ _{S, kt} ^(j) , the output of the GMM 130 at the j-th particle The parameter vector S ^(j) _kt ^(j) _{, t} is obtained from the element distribution k _t ^(j) by sampling shown in the following equation (20).

なお、フレーム選択部２２０はさらに、ＧＭＭサンプリング部２２６に対し、第ｔフレームにおけるＧＭＭの出力パラメータのサンプリングを要求する機能を持つ。

The frame selection unit 220 further has a function of requesting the GMM sampling unit 226 to sample the output parameters of the GMM in the t-th frame.

The updating unit 230 sets the noise parameter in each particle corresponding to the t−1 frame by the extended Kalman filter using the dynamic model composed of the observation equation (1) and the state equation (11) as a state space model. It has a function of updating to the one corresponding to the t frame. At this time, the parameters are updated based on the constraint parameter 138, the state space model 160 (FIG. 3), and the output parameter S ^(j) _kt ^(j) _{, t} sampled by the GMM sampling unit 226. The extended Kalman filter is a Kalman filter corresponding to a state space model including a nonlinear term as shown in the observation equation (1).

FIG. 6 is a block diagram showing the configuration of the update unit 230. Referring to FIG. 6, the updating unit 230 sets the first constraint condition of the state equation (11) for the t−1 frame based on the estimation parameter 205 of the noise probability distribution of the t−1 frame. A weighted average calculation unit 250 for calculating the weighted average vector ^ N _t-1 using the above equation (12) is included.

The updating unit 230 further includes a buffer memory unit 252 for storing the noise parameter for each particle for each frame before the t−1th frame, the noise parameter stored in the buffer memory unit 252, and Based on the number of frames T determined by the constraint condition parameter 138, the Polyak average vector μ _Nt−1 ^(j) for the T frames shown in the above equation (13) in the t−1th frame is calculated for each particle. For the second constraint condition of the state equation (11), based on the Polyak Average calculation unit 254, the Polyak Average vector μ _Nt-1 ^(j), and the estimation parameter 205 of the noise probability distribution in the t−1 frame. A feedback unit 256 for calculating a vector corresponding to the feedback component. The feedback unit 256 calculates the difference μ _Nt-1 ^(j) − ^ N _t-1 ^{(j )} between the Polyak Average vector μ _Nt-1 ^(j) and the average vector ^ N _t-1 ^(j) in the t−1 frame. ⁾ Is calculated.

Further, the updating unit 230 uses the extended Kalman filter whose state space model is the model made up of the observation equation (1) and the state equation (11) to set the noise parameter in the particle corresponding to the t−1 frame to the t And an extended Kalman filter unit 258 for updating to the one corresponding to the frame. The extended Kalman filter unit 258 updates the parameter of the noise in the j-th particle, the observed signal feature amount X _t 124 in the t-th frame, and the output parameter vector S ^(j in the GMM 130 (see FIG. 2) in the j-th particle. ⁾ _kt ^(j) _{, t} , forgetting factor α and scaling factor β given as constraint parameter 138, weighted average vector ^ N _t-1 , and difference μ _Nt-1 ^(j) − ^ N _t-1 ^{( j)} .

Expressions (21) to (26) below show the distribution update formulas of the extended Kalman filter in the present embodiment. In these equations, “ _{t | t−1} ” is attached as a subscript to the parameter in the t-th frame predicted from the parameter corresponding to the t−1 frame.

ただし、行列Σ_Wは、前述したとおり、第ｔ−１フレームから第ｔフレームへの状態変化の際に雑音の特徴量ベクトルＮ_tに生じるランダムガウス雑音ベクトルＷ_t-1の共分散行列を表す。

However, as described above, the matrix Σ _W represents the covariance matrix of the random Gaussian noise vector W _t−1 generated in the noise feature vector N _t when the state changes from the t−1 frame to the t frame. .

Referring to FIG. 5 again, the weight calculation unit 232 includes the feature vector X _t 124 of the observation signal in the t-th frame and the output parameter vector S ^(j) _kt ^(j) of the GMM 130 for each particle in the t-th frame. _{, t} , a mean vector ^ N _t ^(j) and a covariance matrix Σ _Nt ^(j) that are parameters of noise in the particle in the frame, and a weight w _t-1 ^(j) for the particle in the t−1 frame. Based on the above, the weight w _t ^(j) for the particles in the t-th frame is calculated using the calculation methods shown in the above equations (9) and (10). The weights w _t ^(j) (1 ≦ j ≦ J) are normalized so that Σ _{j =} 1 to J w _t ^(j) = 1.

The re-sampling unit 234 has a function of re-sampling the noise parameter of each particle corresponding to the t-th frame according to the noise probability distribution of the particle whose parameter is updated. At this time, the resampling unit 234 does not resample the noise parameter from the probability distribution of noise in the particles to which only a minute weight w _t ^(j) is given. On the other hand, from the probability distribution of particles with a large weight w _t ^(j), resampling is performed a number of times according to the size of the weight w _t ^(j) , and the noise parameters obtained are re-sampled. Allocate the same number of particles as the number of samplings. However, the total number of resampling and the total number of particles are constant (J). This is because the weight assigned to each particle corresponds to the likelihood of the feature quantity X _t 124 of the observation signal as can be seen from the above equation (9).

  The estimation parameter generation unit 236 has a function of regenerating particles corresponding to the t-th frame by the Metropolis-Hastings algorithm of the Markov chain Monte Carlo method. FIG. 7 is a block diagram showing the configuration of the estimation parameter generation unit 236. Referring to FIG. 7, estimated parameter generation section 236 includes a re-update section 262 for re-updating the noise parameter in each particle corresponding to the (t−1) -th frame to that corresponding to the t-th frame. The re-update unit 262 generates a noise probability distribution in the state space model 160 using the noise parameter of each particle obtained by the re-sampling by the re-sampling unit 234. Then, based on the generated probability distribution and the constraint condition parameter 138, the distribution updating formulas shown in the above formulas (21) to (26) are expressed using the same method as the updating unit 230 shown in FIG. The noise parameter of each particle is updated again by the extended Kalman filter.

The estimation parameter generation unit 236 further uses the calculation method shown in the above equations (9) and (10) to calculate the weights for the re-updated particles (hereinafter referred to as “w _t ^{* (j)} ”). A weight recalculation unit 264 for calculation is included.

The estimation parameter generation unit 236 further determines whether or not to allow a re-updated noise parameter from the weight w _t ^(j) for the re-sampled particle and the weight w _t ^{* (j)} for the re-updated particle. An allowable probability calculating unit 266 for calculating an allowable probability ν used for the determination, a random number generating unit 268 for generating a random number u within a closed interval from 0 to 1 by a predetermined random number generating method, an allowable probability ν, A parameter selection unit 270 for selecting one of the noise parameter of the resampled particle and the noise parameter of the reupdated particle as the parameter of the particle corresponding to the t-th frame based on the random number u; including.

The allowable probability calculation unit 266 has a function of calculating the allowable probability ν from the weight w _t ^(j) and the weight w _t ^{* (j) according} to the following equation (27).

パラメータ選択部２７０は、乱数ｕが許容確率ν以下であれば、当該パーティクルにおける雑音のパラメータ及びその重みを再更新で得られた新たなパラメータ及びその重みに変更する機能を持つ。

If the random number u is less than or equal to the allowable probability ν, the parameter selection unit 270 has a function of changing the noise parameter and its weight in the particle to a new parameter and its weight obtained by re-update.

<Realization by computer>
As will be apparent from the following description, the preprocessing unit 104, the preprocessing acoustic model unit 106, and the search unit 110 of the speech recognition system 100 shown in FIG. 1 are all executed on computer hardware. And a program stored in computer hardware. FIG. 8 is a flowchart showing a control structure of a computer program that realizes noise suppression processing performed by the noise suppression unit 114 included in the preprocessing unit 104 (see FIG. 1).

Referring to FIG. 8, when the noise suppression process is started, in step 282, an initial distribution corresponding to the value of each element of noise feature amount N ₀ in the initial state is estimated. That is, the average vector μ _N and the covariance matrix Σ _N that are parameters of the initial noise distribution p (N ₀ ) shown in the equation (4) are calculated by the calculation methods shown in the equations (15) and (16). . Further, the vector N ₀ ^(j) (j = 1,..., J) is sampled from the noise initial distribution p (N ₀ ) according to the equations (17) and (18), and the noise parameters for each particle in the initial frame are sampled. Is estimated. In addition the step 282, is substituted for L-1 to the variable _{I L.} That is, the calculation control unit 212 illustrated in FIG. 4 stores L−1 in the number storage unit 218.

  In step 284, the frame subject to noise suppression is shifted to the next frame. In the following description, it is assumed that the frame after the transition is the t-th frame.

In step 285, the noise parameter in each particle is estimated for the processing target frame using the particle filter. That is, the mean vector ^ N _t ^(j) and the covariance matrix Σ _Nt ^(j) that are parameters of the probability density function p (N _{0: t} ^(j) | X _{0: t} ⁾ are estimated, and further, for each particle. A weight w _t ^(j) is determined, and an estimation parameter 205 of the noise probability distribution is generated. The processing in this step will be described later with reference to FIG.

In step 286, 1 is added to the variable _{I L.}

In step 288, the value of the variable I _L is determined whether equal to the value of the constant L, control flow depending on the result of the determination is branched. That is, if this determination result is YES, control proceeds to the subsequent step 292, and otherwise, control proceeds to step 298 described later.

In step 292, the observed signal distribution parameter 208 is estimated. That is, the probability distribution of the feature quantity X _t 124 of the observed signal in each particle is estimated using the noise parameters ^ N _t ^(j) and Σ _Nt ^(j) determined in step 285. Further, for each element distribution k (1 ≦ k ≦ K) constituting the GMM 130, the average vector μ _Xkt ^(j) _{, t} of the probability distribution of the observed signal feature quantity X _t 124 in the particle and the covariance matrix Σ _{Xk, t} ^(j) is calculated.

In step 294, the average vector and covariance matrix of the probability distribution of the observed signal estimated in step 292 are stored in the distribution storage unit 214 (see FIG. 4). The value of the subsequent step 296 variable _{I L} is cleared to 0. Thereafter, control proceeds to step 300.

  On the other hand, if the determination result in step 288 is NO, the control proceeds to step 298. In step 298, the average vector and covariance matrix stored in the distribution storage unit 214 when the determination result in step 288 is YES are read from the distribution storage unit 214. After step 298, control proceeds to step 300.

In step 300, the feature quantity 126 of the estimated clean speech in the t-th frame is calculated by the MMSE estimation method. That is, first, the MMSE estimation value vector Ｓ _St is calculated by the MMSE estimation method using the parameters obtained by the processing of step 285 and step 292 or the parameters read from the distribution storage unit 214 in step 298, and the estimation is performed. Output as clean speech feature 126 (see FIG. 1).

この式において、Ｐ（ｋ｜Ｘ_t，（ｊ））は、ｊ番目のパーティクルにおける、ＧＭＭ１３０内の要素分布ｋに対する混合重みを表す。混合重みＰ（ｋ｜Ｘ_t，（ｊ））は、特許文献１に記載されたものと同様、次の数式により算出される。

In this equation, P (k | X _t , (j)) represents the mixing weight for the element distribution k in the GMM 130 in the j-th particle. The mixing weight P (k | X _t , (j)) is calculated by the following equation, as described in Patent Document 1.

ただし、式（２８）〜（３１）はｔ＝ｎ・Ｌ（ｎはゼロ又は正の整数）のときの式である。ｔ≠ｎ・Ｌのときには、観測信号分布推定部２０２で推定されたパラメータではなく、分布記憶部２１４に記憶されたパラメータを用いるので、この式は次のとおりとなる。

However, Expressions (28) to (31) are expressions when t = n · L (n is zero or a positive integer). When t ≠ n · L, since the parameter stored in the distribution storage unit 214 is used instead of the parameter estimated by the observation signal distribution estimation unit 202, this equation is as follows.

この式（３２）で、ｔ_Ｌは、ｔ_Ｌ＝（ｎ−１）・Ｌ＜ｔ＜ｎ・Ｌ（ｎは正の整数）となる条件を満たす値である。

In this equation (32), t _L is a value that satisfies the condition of t _L = (n−1) · L <t <n · L (n is a positive integer).

This will be described with reference to FIG. FIG. 9 shows the relationship between the frame time (t) when L = 4 and the parameters of the observed signal distribution used for calculating the mixing weight in each frame. Referring to FIG. 9, it is assumed that the parameters of the observation signal distributed at a certain time t = t ₀ is calculated. Thereafter, in the frames t = t ₁ , t _2, and t ₃ until the parameter of the observed signal distribution is calculated again at t = t ₄ after 4 frames, the parameter of the observed signal distribution calculated at t = t ₀ Is used for weight calculation.

In step 300, the MMSE estimation value vector ^ _St is calculated according to the equations (28) to (30) by the MMSE estimation method using the mixture weight calculated in this way, and the estimated clean speech feature 126 is obtained. (See FIG. 1).

  Subsequently, in step 302, end determination is performed. That is, if the t-th frame is the final frame, the noise suppression process is terminated. Otherwise return to step 284.

  In this way, the observation signal distribution estimation unit 202 estimates the observation signal distribution every L frames, but not in the frames in between. As a result, the processing time in the observation signal distribution estimation unit 202 can be shortened. As a result, the calculation amount of this part is 1 / L simply considering, and the calculation amount for the estimation of clean speech can be greatly reduced.

  FIG. 10 is a diagram showing in which frame the observation signal distribution is estimated when the value of L is 1, 2, 4, and 8. In FIG. When L = 1, estimation is performed for all frames. On the other hand, if L = 2, the number of frames to be estimated is ½ as apparent from FIG. When L = 4 and L = 8, they are 1/4 and 1/8, respectively.

FIG. 11 is a flowchart showing a control structure of a program that realizes the generation process of the estimation parameter 205 of the noise probability distribution performed in step 285 (see FIG. 8). Referring to FIG. 11, when the generation process of the estimation parameter of the noise probability distribution is started, in step 320, the first and second constraint conditions for the state transition process of noise 121 when updating by the extended Kalman filter are set. Such a parameter vector is calculated. That is, the weighted average vector ^ N _t-1 of the noise parameter at the particle of the t−1th frame is calculated using Expression (12). Then, for each particle, a Polyak Average vector μ _Nt−1 ^(j) is calculated from the noise parameters of the particle for the past T frames, and a difference μ _Nt−1 from the average vector ^ N _t−1 ^(j) is calculated. ^(j) − ^ N _t−1 ^(j) is calculated.

  In step 322, the noise parameter in each particle in the t-th frame is estimated from the noise probability distribution in the particle in the t-1 frame using the extended Kalman filter expressed by the equations (21) to (26).

In step 324, the weight w _t ^(j) for each particle in the t-th frame is calculated by the equations (9) and (10). Then, the weight w _t ^(j) is normalized. In step 326, the number of re-sampling from each particle is determined based on the weight w _t ^(j) for each particle, and the parameter is re-sampled based on the noise probability distribution in the particle. In step 328, the particles of the t-th frame are regenerated using the Metropolis-Hastings algorithm.

FIG. 12 is a flowchart showing details of processing in step 328 (see FIG. 11). Referring to FIG. 12, when the processing in step 328 is started, in step 340, as in step 320 shown in FIG. 11, the weighted average vector ^ N _t-1 is calculated by the calculation method shown in equation (12). calculate. Then, for each particle, a Polyak Average vector μ _Nt−1 ^(j) is calculated from the noise parameters of the particle for the past T frames, and a difference μ _Nt−1 from the average vector ^ N _t−1 ^(j) is calculated. ^(j) − ^ N _t−1 ^(j) is calculated.

In the subsequent step 342, using the noise probability distribution expressed by the noise parameter in each particle obtained by re-sampling in step 326 (see FIG. 11), the extended Kalman filter shown in equations (21) to (26) is used. The noise parameters in each particle are updated again. In other words, particles at the t-th frame are newly prepared, and corresponding to the particles at the t-th frame from the parameters corresponding to the particles at the (t-1) -th frame by the same process as the process at step 322 (see FIG. 11). Update the parameters again and set the parameters of the prepared particles. In step 344, the weight w _t ^{* (j)} for the particles prepared in step 342 is calculated and normalized by the same process as the process in step 324 shown in FIG.

In step 346, the allowable probability ν of the particles prepared in step 342 is determined by comparing the weight w _t ^(j) calculated in step 324 with the weight w _t ^{* (j)} calculated in step 344. Determine. In step 348, a random number u is generated by selecting an arbitrary value from the uniform set U _[0,1] consisting of values in the interval [0,1]. In step 350, the value of the random number u generated in step 348 is compared with the value of the allowable probability ν determined in step 346. If u is less than or equal to the allowable probability, the process proceeds to step 352. Otherwise, go to step 354. In step 352, the particles prepared in step 342 are allowed. That is, the parameter obtained by the resampling in step 326 is replaced with the parameter of the prepared particle, and the process is terminated. In step 354, the particles prepared in step 342 are rejected. That is, the prepared particles and their parameters are rejected, and the process ends.

[Operation]
The speech recognition system 100 according to the present embodiment operates as follows.

First, an operation in which the noise probability distribution estimation unit 200 shown in FIG. 5 generates the noise probability distribution estimation parameter 205 in the initial frame (t = 0) will be described. The measurement unit 112 illustrated in FIG. 1 receives the noise-superimposed speech 122 from the sound source 102, and extracts the feature amount X _t 124 of the observation signal. The extracted feature amount X _t 124 is given to the noise probability distribution estimation unit 200 shown in FIG. 5 of the noise suppression unit 114.

With reference to FIG. 5, the frame selection unit 220 of the noise probability distribution estimation unit 200 gives the first 10 frames of the feature amount X _t 124 to the noise initial distribution estimation unit 222. The initial noise distribution estimation unit 222 estimates the initial noise distribution p (N ₀ ) by the processing shown in the above equations (14) to (16). Further, sampling shown in the above equations (17) and (18) is performed J times from the initial noise distribution p (N ₀ ). By this sampling, a vector N ₀ ^(j) and a covariance matrix Σ _N0 ^{(j), which} are initial parameters of noise in each particle, are determined. The noise probability distribution estimation unit 200 outputs these parameters as an estimation parameter 205 of the noise probability distribution in the initial frame.

  A value L is preset in the interval storage unit 206 illustrated in FIG. 4, and a value L−1 is stored in the number storage unit 218.

Next, an operation in which the successive estimation unit 224 of the noise probability distribution estimation unit 200 generates the noise probability distribution estimation parameter 205 in the t-th frame (t ≧ 1) will be described. Referring to FIG. 5, in response to processing start request 210 for the next frame, frame selection unit 220 requests GMM sampling unit 226 to sample the output parameter of GMM in the t-th frame and observe signal It gives the feature quantity X _t 124 of the updating section 230.

The GMM sampling unit 226 samples the output parameter vector S ^(j) _kt ^(j) _{, t} from the GMM 130. For example, at the j-th particle, the GMM sampling unit 226 samples the element distribution k _t ^(j) with a probability according to the mixing weight from the mixed normal distribution 140 in the GMM 130 shown in FIG. As a result, the element distribution 150 is sampled as the element distribution k _t ^(j) . The GMM sampling unit 226 further samples the output parameter vector S ^(j) _kt ^(j) _{, t} according to the output probability distribution represented by the element distribution k _t ^(j) . The GMM sampling unit 226 samples the output parameter vectors S ^(j) _kt ^(j) _{and t} for the total number J of particles in accordance with the above-described procedure, and supplies the sampled data to the updating unit 230 shown in FIG.

FIG. 13 schematically shows an outline of parameter updating and re-sampling performed by the sequential calculation unit 224. In FIG. 13, a certain noise parameter is distributed in the left-right direction, and the time advances from top to bottom. Further, in FIG. 13, the particles are schematically shown by white circles and black circles. For example, a particle indicated by a white circle is a minute particle having a value of weight w _t ^(j) , and a particle indicated by a black circle is a particle having a large value of weight w _t ^(j) To do.

  Referring to FIG. 13, it is assumed that state space 420 is approximately represented by particles corresponding to the (t-1) th frame. The updating unit 230 updates the noise parameter of each particle in the state space 420 to the noise parameter of each particle in the state space 430 corresponding to the t-th frame as follows.

  First, the extended Kalman filter unit 258 of the updating unit 230 illustrated in FIG. 6 acquires the estimation parameter 205 of the estimated probability distribution of each particle in the t−1 frame. The obtained estimation parameter 205 of the estimated probability distribution is given to the weighted average calculation unit 250, the buffer memory 252, and the feedback unit 256. At this time, the buffer memory 252 stores the estimated parameter 205 of the estimated probability distribution for at least T frames before the (t-1) th frame.

The weighted average calculation unit 250 shown in FIG. 6 calculates the weighted average vector ^ N _t-1 shown in Expression (12) when the estimation parameter 205 of the estimated probability distribution is given. When the first constraint condition in the state equation shown in the equation (11) is introduced based on the weighted average vector ^ N _t−1 and the noise average vector is corrected, the noise parameter in the corrected noise probability distribution is The weighted average vector ^ N _t-1 is closer than the average vector ^ N _t-1 ^(j) before correction. Therefore, scattering of particles is suppressed.

When the estimated parameter 205 of the new estimated probability distribution is accumulated in the buffer memory unit 252, the Polyak Average calculating unit 254 uses the estimated parameter 205 of the estimated probability distribution for T frames accumulated in the buffer memory unit 252. Then, the Polyak Average vector μ _Nt ^(j) shown in Expression (13) for each particle is calculated. The calculated Polyak Average vector μ _Nt−1 ^(j) is given to the feedback unit 256. Feedback unit 256, in each particle, and Polyak Average, which vector μ _Nt-1 ^(j), the average vector ^ N _t-1 ^(j) the difference between ^{_{μ Nt-1 (j) -}} ^ N t-1 (j) Is calculated. When the estimation parameter 205 of the estimated probability distribution is not accumulated in the buffer memory unit 252 for T frames, the Polyak Average calculation unit 254 estimates the noise probability distribution for only the frames accumulated in the buffer memory unit 252. Using the parameter 205, the Polyak Average vector μ _Nt ^(j) is calculated.

FIG. 14 schematically shows the concept of Polyak Average and feedback. 14A and 14B both show the Polyak Average vector μ _Nt ^(j) in the j-th particle and the noise feature vector N _t-4 ^(j) ,..., N _{t +} corresponding to the particle. ₁ represents the relationship with ^(j) . FIG. 14A shows a case where the time transition of the noise feature vector is gentle, and FIG. 14B shows a case where the time transition is intense. In these figures, time progresses from left to right, and the feature amount of noise changes in the vertical direction. In FIGS. 14A and 14B, the Polyak Average vector μ _Nt ^(j) in the t-th frame is indicated by a white circle. In the Polyak Average vector μ _Nt ^(j) shown in this figure, it is assumed that T = 5 frames.

Referring to FIG. 14A, the noise feature amount N _t-1 ^{(j) in the (} _t−1 ^{) th} frame and the Polyak
A difference μ _Nt ^(j) −N _t ^(j) is generated between the Average vector μ _Nt ^(j) . Similarly, in the case of FIG. 14 (B) intense such time transition as shown, the noise characteristic amount N _t ^(j), between the Polyak Average, which vector mu _Nt ^(j), the difference mu _Nt ^(j) −N _t ^(j) is generated. Figure 14 (A) of noise in the feature vector ^{_{N t-4 (j),}} ..., N t as compared with the variation of the ^(j), FIG. 14 (B) the noise of the feature in the vector N _t-4 ^(j) , ..., N _t ^(j) varies greatly. That is, the difference between the noise feature vectors N _t-4 ^(j) ,..., N _t ^(j) in FIG. 14 (A) is smaller than those differences in FIG.

The Polyak Average vector μ _Nt ^(j) is the average of N _t−4 ^(j) ,..., N _t ^(j) . Therefore, the possible range of the Polyak Average vector μ _Nt ^(j) is the range from the minimum to the maximum of N _t−4 ^(j) ,..., N _t ^(j) . Therefore, as shown in FIG. 14A, if the difference between these feature quantity vectors is small, the range that the Polyak Average vector μ _Nt−1 ^(j) can take is narrowed accordingly. The fluctuation range of the difference μ _Nt−1 ^(j) −N _t−1 ^(j) is naturally reduced. On the other hand, as shown in FIG. 14B, if the difference between the noise feature vectors is large, the range that the Polyak Average vector μ _Nt ^(j) can take is widened accordingly. The fluctuation range of the difference μ _Nt ^(j) −N _t ^(j) naturally increases. That is, the difference μ _Nt ^(j) −N _t ^(j) reflects the noise change for the past T frames. Based on this difference, when the feature vector N _{t + 1} ^(j) of noise in the next frame is predicted, a feature vector reflecting the noise change for the past T frames is obtained.

The extended Kalman filter unit 258 (see FIG. 6) includes a weighted average vector ^ N _t−1 , a difference vector μ _Nt−1 ^(j) −N _t−1 ^(j), and a forgetting factor α determined by a constraint parameter 138. Each particle is updated by the extended Kalman filter expressed by the equations (21) to (26) based on the scaling coefficient β, the observed signal feature amount X _t 124, and the output parameter 240.

In this update, the dispersion of ^ N _t-1 ^(j) is suppressed in the one-period ahead prediction parameter N _{t | t-1} ^(j) of noise shown in Expression (21). Also, parameter variations for the past T frames are fed back. That is, when the past fluctuation is large, the fluctuation of the one-year ahead prediction parameter N _{t | t−1} ^(j) also becomes large. On the other hand, when the past fluctuation is small, the fluctuation of the one-year prediction parameter N _{t | t−1} ^(j) is also small. Therefore, the constraint condition for the time transition of the parameter is strengthened by the past parameter variation.

  By updating each particle as described above, each particle in the state space 420 shown in FIG. 13 is updated, and the state space 430 corresponding to the t-th frame is expressed by the particle whose parameter is updated. The

In response to this, the weight calculation unit 232 calculates the weight w _t ^(j) for each particle in the state space 430 by the equations (22) and (23). The re-sampling unit 234 re-samples the noise parameter in the particle based on the weight w _t ^(j) . At this time, the resampling unit 234 first sets the number of resamplings from each particle in the state space 430 for each particle according to the weight w _t ^(j) for the particle. The number of samplings from a minute particle with a weight represented by a white circle is set to zero. In addition, the number of times of sampling from particles with a large weight represented by black circles is set to 1 to 3 according to the magnitude of the weight. Subsequently, based on the noise probability distribution of the particles in the state space 430, the noise parameters are resampled by the set number of times. In this way, particles representing a new state space 440 corresponding to the t-th frame are formed.

When such re-sampling by the re-sampling unit 234 is repeatedly performed, the noise parameter in many particles corresponding to a certain frame is obtained from the probability distribution of the noise parameter in a small number of particles corresponding to the previous frame. May be sampled. Therefore, the estimated parameter generation unit 236 prevents such a situation by newly generating parameters for the particles corresponding to the t-th frame using the Metropolis-Hastings algorithm. The re-updating unit 262 illustrated in FIG. 7 re-updates the noise parameters of the particles in the state space 420 corresponding to the (t-1) th frame, according to the noise probability distribution in the state space 440. The weight recalculation unit 264 calculates a weight w _t ^{* (j)} for the re-updated particle. Acceptable probability calculation unit 266, the weight w _t ^* for particles that are re-updated ^(j), based on the weight w _t ^(j) with respect to the resampled particles, calculates the permission probability [nu. The parameter selection unit 270 compares the allowable probability ν with the random number u in the interval [0, 1] generated by the random number generation unit 268. If the random number u is equal to or less than the allowable probability ν, the parameter selection unit 270 Replace the parameter with the parameter in the re-updated particle. Otherwise, reject the parameter in the re-updated particle.

By repeating the above operation for each frame, the mean vector ^ N _t ^(j) and the covariance matrix Σ _Nt ^(j) , which are noise parameters for each particle, are estimated corresponding to each frame. . The average vector ^ N _t ^(j) and the covariance matrix Σ _Nt ^(j) , which are noise parameters for each particle, and the weight w _t ^(j) for each particle are the estimation parameters 205 of the noise probability distribution. The noise probability distribution estimation unit 200 supplies the noise probability distribution estimation parameter 205 and the observation signal feature vector X _t 124 to the calculation control unit 212 and the observation signal distribution estimation unit 202 shown in FIG. 4 for each frame.

Referring to FIG. 4, the calculation control unit 212 adds 1 to the value stored I _L in the frequency storage unit 218, if the value I _L is equal to the value L of the stored frame interval to the interval storage unit 206 Determine whether or not.

When the value IL is equal to the value _L of the frame interval, the calculation control unit 212 gives an estimation instruction signal 209 to the observation signal distribution estimation unit 202, and sends an observation signal distribution storage control signal 207 to the distribution storage unit 214. give. Always value I _L is equal to the value L in the first process. Therefore, for the first frame, the observation signal distribution parameter 208 by the observation signal distribution estimation unit 202 is estimated and stored in the distribution storage unit 214.

  When the estimation instruction signal 209 is given from the calculation control unit 212, the observation signal distribution estimation unit 202 shown in FIG. 4 is based on the noise probability distribution estimation parameter 205 and the GMM 130 as the observation signal distribution parameter 208 by the VTS method. Then, an average vector and a covariance matrix of the observation signal distribution in each particle corresponding to the t-th frame are generated. As a result, the probability distribution of noise and the probability distribution of the observation signal are estimated for each particle. These values are given to the selection unit 216 and the distribution storage unit 214.

  The distribution storage unit 214 stores an observation signal distribution parameter 208 in response to the observation signal distribution storage control signal 207.

  The calculation control unit 212 sets the value of the selection instruction signal 211 as the first value. Since the selection instruction signal 211 is the first value, the selection unit 216 selects the observation signal distribution parameter 208 that is the output of the observation signal distribution estimation unit 202 and uses the observed signal distribution parameter 217 as the clean speech estimation unit 204. To give.

The clean speech estimation unit 204 calculates the MMSE estimated value vector ^{ circumflex over ⁽ S ⁾ _} ⁽ _t ⁾ of the clean speech 120 for each particle corresponding to the t-th frame by the MMSE estimation method. Further, using the MMSE estimated value vector ^{ circumflex over ⁽ S ⁾ } _t ^(j) and the weight w _t ^(j) , the estimated clean speech feature vector ^{ circumflex over ⁽ _t ⁾ _} 126 in the t-th frame is calculated, and the search unit 110 shown in FIG. Output to.

When the value I _L is not equal to the value L of the frame interval, calculation control unit 212 does not give an estimate indication signal 209 with respect to the observed signal distribution estimation unit 202. Therefore, the observation signal distribution estimation unit 202 does not estimate the observation signal distribution. At this time, the calculation control unit 212 does not give the observation signal distribution storage control signal 207 to the distribution storage unit 214. As a result, the distribution storage unit 214 continues to store the observation signal distribution finally estimated by the observation signal distribution estimation unit 202. The calculation control unit 212 further sets the value of the selection instruction signal 211 as the second value. Since the selection instruction signal 211 has the second value, the selection unit 216 selects the parameter 213 of the observation signal distribution estimated last, which is the output of the distribution storage unit 214, and clean speech as the observation signal distribution parameter 217. This is given to the estimation unit 204.

Therefore, the clean speech estimation unit 204 uses the observed signal distribution parameter 208 calculated for the t _L frame (t _L <t) for each particle corresponding to the t frame by the MMSE estimation method. 120 MMSE estimated value vectors ^{ circumflex over ⁽ S ⁾ } _t ^(j) are calculated. Further, using the MMSE estimated value vector ^{ circumflex over ⁽ S ⁾ } _t ^(j) and the weight w _t ^(j) , the estimated clean speech feature vector ^{ circumflex over ⁽ _t ⁾ _} 126 in the t-th frame is calculated, and the search unit 110 shown in FIG. Output to.

The search unit 110 shown in FIG. 1 uses the estimated clean speech feature vector ^ _St 126 to obtain the acoustic model held in the recognition acoustic model unit 109 and the language model held in the language model unit 108. Based on this, a word or the like of a target language that matches is searched, and the result is output as a recognition output 128.

  As described above, according to the speech recognition system 100 according to the present embodiment, the amount of calculation in the estimation of the observation signal distribution can be reduced to 1 / L by changing the value of L. In the calculation amount of the entire noise removal by the particle filter, the calculation amount in the estimation of the observed signal distribution is about half of the whole. Therefore, if the calculation amount related to the estimation of the observation signal distribution can be reduced as in the above embodiment, the calculation amount of the entire noise removal can be greatly reduced, and as a result, the noise removal can be speeded up. Alternatively, it is possible to achieve noise removal with equivalent performance using less computational resources.

[Voice recognition experiment]
In order to evaluate the effectiveness of noise removal shown in the above embodiment, a continuous speech recognition experiment was conducted. As the evaluation voice, the 510 utterance voice extracted as testset-01 from the utterance corpus called BTEC prepared by the applicant was used. This uttered voice is a clean voice, but the noise recorded at the station concourse and the noise recorded near the station ticket gate were superimposed on this voice under the conditions of 25, 15, and 10 dB, respectively. The acoustic model used in the evaluation experiment was learned using TRA, TRA-BLA, APP-BLA (about 170,000 utterances) created by the applicant.

  The experimental results are shown in Table 1 below. The amount of calculation shown in Table 1 is a value calculated by a computer having a CPU (Central Processing Unit) having a normal performance and operating commercially with a clock signal of 1.2 GHz. In Table 1, for example, “0.64 × RT” means that it takes 6.4 seconds to process 10-second noise-superimposed speech.

テーブル１の結果から明らかなように、観測信号分布の推定を間引くことにより、推定間隔Ｌ＝１６のときには、従来の推定間隔Ｌ＝１の場合と比較して、約２８％の計算時間が削減されている。しかも、計算時間が大幅に削減されているにもかかわらず、単語正解精度により表される音声認識性能はほとんど同等である。

As is clear from the results of Table 1, the calculation time is reduced by about 28% when the estimation interval L = 16 by thinning out the estimation of the observed signal distribution compared to the conventional estimation interval L = 1. Has been. In addition, the speech recognition performance expressed by the word correct accuracy is almost the same, although the calculation time is greatly reduced.

  From this experimental result, it can be seen that according to the present embodiment, it is possible to maintain good speech recognition performance while reducing the amount of calculation in an environment where non-stationary noise exists.

[Modification]
In the above-described embodiment, a Gaussian mixture distribution prepared in advance by statistical processing (learning) for sample speech is used as the clean acoustic model. The Gaussian mixture distribution is a multi-dimensional distribution including a plurality of element distributions for each dimension. The average and variance are calculated for each element distribution by prior learning. Therefore, in many cases, the mean and variance are different for each element distribution. Therefore, even a complex distribution can be statistically modeled. In this case, it is possible that the averages of the two element distributions match or the variances of the two element distributions match, but they do not usually match.

  However, in order to realize noise suppression according to the present invention, the element distribution obtained by learning need not be used as it is. For example, in the above-described embodiment, noise suppression can be performed using the same mechanism as in the above-described embodiment even if only the average of the element distribution is used and the variance is assumed to be the same in all element distributions. it can. In this case, only the average of each element distribution may be stored as the acoustic model.

  Furthermore, when calculating the average of the element distribution, the feature amount may be calculated as a continuous value, or the feature amount is set to a discrete value in advance, and the feature amount obtained by the calculation is the closest discrete Quantization may be performed by replacing the target feature amount.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is the schematic which shows the structure of the speech recognition system 100 which concerns on one embodiment of this invention. It is the schematic which shows the concept of GMM130. It is the schematic which shows the concept of the state space model 160 of an observation signal. 3 is a block diagram illustrating a configuration of a noise suppression unit 114. FIG. 3 is a block diagram illustrating a configuration of a noise probability distribution estimation unit 200. FIG. 3 is a block diagram illustrating a configuration of an update unit 230. FIG. 4 is a block diagram illustrating a configuration of an estimation parameter generation unit 236. FIG. It is a flowchart which shows the control structure of a noise suppression process. It is a schematic diagram which shows the relationship between the parameters of the observed signal distribution used in MMSE estimation when L = 4. It is a schematic diagram which shows the timing of parameter estimation of the observation signal distribution when L = 1, 2, 4, 8. It is a flowchart which shows the control structure of the production | generation process of the estimation parameter 205 of noise probability distribution. It is a flowchart which shows the control structure of the sampling process by a Metropolis-Hastings algorithm. It is a figure which shows the outline | summary of the process by a particle filter. It is a schematic diagram which shows the concept of Polyak Average and feedback. FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Speech recognition system 102 Sound source 104 Preprocessing part 106 Preprocessing acoustic model part 108 Language model part 109 Recognition acoustic model part 110 Search part 112 Measurement part 114 Noise suppression part 116 Speaker 118 Noise source 120 Clean voice 121 Noise 122 Noise Superimposed speech 124 Observed signal feature 126 Estimated clean speech feature 130 GMM
132 learning data storage unit 134 model learning unit 136 GMM storage unit 138 constraint parameter 160 state space model 200 noise probability distribution estimation unit 202 observation signal distribution estimation unit 204 clean speech estimation unit 206 interval storage unit 208 observation signal distribution parameter 214 distribution Storage unit 216 Selection unit 218 Number storage unit 220 Frame selection unit 222 Noise initial distribution estimation unit 224 Sequential calculation unit 226 GMM sampling unit 230 Update unit 232 Weight calculation unit 234 Re-sampling unit 236 Estimation parameter generation unit 240 Output parameter 250 Weighted average calculation Unit 252 buffer memory unit 254 Polyak Average calculation unit 256 feedback unit 258 extended Kalman filter unit 262 re-update unit 264 weight recalculation unit 266 allowable probability calculation unit 268 random number generation 270 parameter selection unit

Claims (8)

A noise suppression device for suppressing a noise component in an observation signal obtained by observation of a target voice in an environment where noise is generated,
The observation signal is made up of a plurality of element distributions prepared in advance using a particle filter having a plurality of particles, each receiving a feature amount extracted from a frame having a predetermined time length framed at predetermined intervals. Noise estimation means for sequentially generating, for each frame, an estimation parameter of a probability distribution representing the noise based on an acoustic model for clean speech estimation;
Adapting means for adapting the acoustic model to noise according to a noise probability distribution whose parameters are estimated by the noise estimating means;
Target speech estimation means for calculating an estimated feature amount of the target speech for each frame by the MMSE estimation method using the acoustic model and the feature amount of the observation signal;
A noise suppression device including control means for controlling an interval of adaptation by the adaptation means so that the adaptation of the acoustic model by the adaptation means is performed for each of a plurality of frames,
The target speech estimation means calculates an estimated feature amount of the target speech for the frame using the adapted acoustic model when the adaptation means performs adaptation on the frame. When adaptation by the adapting means is not performed for a certain frame, the acoustic model adapted for the previous frame by the adapting means is used for the previous frame. A noise suppression device characterized by calculating an estimated feature amount of a target speech. The noise suppression device according to claim 1, wherein the acoustic model is a Gaussian mixture distribution including the plurality of element distributions. The noise suppression device according to claim 2, wherein variances of any two of the plurality of element distributions are different from each other. The noise suppression device according to claim 2, wherein the plurality of element distributions have equal variances. And further comprising storage means for storing the adapted acoustic model in response to the adaptation of the acoustic model for the clean speech estimation for a frame by the adaptation means,
When the acoustic model for the clean speech estimation is adapted for a certain frame, the target speech estimation means uses the adapted acoustic model and estimates the target speech by the MMSE estimation method. When the acoustic model is not adapted for the clean speech estimation for a certain frame, the adaptive acoustic model stored in the storage means is used to perform the MMSE method. The noise suppression apparatus according to claim 1, wherein an estimated feature amount of the target speech is calculated. The control means includes
Interval storage means for storing in advance information for determining an interval between frames for performing processing for adapting the acoustic model for the clean speech estimation to noise,
Frame number storage means for storing the number of frames processed immediately after the adaptation by the adaptation means is performed immediately before;
Adding means for adding 1 to the stored content of the frame number storage means each time a frame to be processed is given to the noise suppression device;
A determination means for determining whether or not the storage content of the frame number storage means and the storage content of the interval storage means are equal;
Means for performing processing for enabling adaptation to the frame by the adaptation means, and processing for disabling adaptation to the frame by the adaptation means, according to a determination result by the determination means;
Means for clearing the frame number storage means to zero in response to the determination means determining that the storage content of the frame number storage means and the storage content of the interval storage means are equal. The noise suppression device according to any one of claims 1 to 5. A computer program that, when executed by a computer, causes the computer to operate as the noise suppression device according to any one of claims 1 to 6. The noise suppression device according to any one of claims 1 to 6,
Voice recognition means for receiving the estimated feature quantity of the target voice calculated by the noise suppression device and performing voice recognition on the target voice using the acoustic model and a predetermined language model on the recognition target language And a voice recognition system.

Images