JP6243858B2 - Speech model learning method, noise suppression method, speech model learning device, noise suppression device, speech model learning program, and noise suppression program - Google Patents

Description

  The present invention relates to a speech model learning method, a noise suppression method, a speech model learning device, a noise suppression device, a speech model learning program, and a noise suppression program.

  In recent years, the use of automatic speech recognition has been increasing in the information society, and technological advances are highly expected. When automatic speech recognition is used in an actual environment, it is necessary to remove noise from a signal other than a speech signal to be processed, that is, an acoustic signal including noise, and extract a desired speech signal.

  For example, a signal obtained by mixing a speech signal and a noise signal is input, and a probability model of the input mixed signal is generated from the probability models of the speech signal and the noise signal estimated in advance. At that time, the difference between the probability model of each of the speech signal and the noise signal constituting the probability model of the input mixed signal and the statistics of each of the speech signal and the noise signal included in the input mixed signal is expressed by Taylor series approximation. The difference is estimated using the EM algorithm, and the input mixed signal probability model is optimized. Thereafter, a method of suppressing noise using the optimized probabilistic model of the input mixed signal and the parameters of the probabilistic model of the speech signal is disclosed (for example, see Non-Patent Document 1).

  In addition, for example, a signal obtained by mixing a speech signal and a noise signal is input, and a probability model of a speech signal learned by using speech data for learning of a large number of speakers is used for the speaker of the speech signal included in the input mixture signal. In order to adapt to the feature (speaker adaptation) and deal with a noise signal whose statistical properties follow a multimodal distribution, a speech signal and a noise signal are extracted from the input mixed signal, respectively. At this time, the reliability of each extracted signal is calculated for each unit time with the SN ratio as a reference. The extracted speech signal and noise signal and the reliability of each signal are used to estimate a speaker adaptation parameter and a stochastic model of the noise signal according to a multimodal distribution using an EM algorithm. Then, the optimal probability model of the input signal is generated from the probability model of the speech signal after speaker adaptation and the estimated probability model of noise, and the optimal probability model of the input mixed signal and the speech signal after speaker adaptation are generated. A method of suppressing noise using a parameter of a probability model is disclosed (for example, see Non-Patent Document 2).

P. J. Moreno, B. Raj, and R. M. Stern, “A vector Taylor series approach for environment-independent speech recognition.” In Proceedings of ICASSP '96, vol. II, pp. 733-736, May 1996. M. Fujimoto and T. Nakatani, “A reliable data selection for model-based noise suppression using unsupervised joint speaker adaptation and noise model estimation.” In Proceedings of ICSPCC '12, pp. 4713-4716, Aug 2012.

  However, the above prior art is based on the assumption that, for example, in Non-Patent Document 1, the characteristics of the noise signal included in the input mixed signal are stationary and the distribution (frequency distribution or probability distribution) is unimodal. This is a technology for noise suppression. However, many noise signals in the real environment have non-stationary characteristics, and their distribution is often multimodal. Therefore, it cannot cope with non-stationary noise signals and sufficient noise suppression performance cannot be obtained. In addition, since the relationship between the speech signal and the noise signal included in the input mixed signal is expressed by a nonlinear function, an analytical solution can be used when estimating the parameters of the probability model of the speech signal and the noise signal even if Taylor series approximation is used. I can't get it. Therefore, the optimal solution of the probability model parameter for each of the speech signal and the noise signal cannot be obtained, and sufficient noise suppression performance cannot be obtained.

  Further, in the non-patent document 2, for example, the above prior art estimates a noise signal probabilistic model according to a multimodal distribution, so that it is possible to handle a non-stationary noise signal. An adaptation parameter and a stochastic model of a noise signal following a multimodal distribution are estimated by an EM algorithm. A Gaussian Mixture Model (GMM) is used for the probabilistic model of the speech signal. When the speech signal and the noise signal are extracted from the input mixture signal and the speaker adaptive parameter estimation and the noise suppression filter are designed. Therefore, posterior probabilities (defined as speech posterior probabilities) for each element distribution included in the GMM of the speech signal are required. This corresponds to an identification problem as to which element distribution in the GMM of the audio signal the audio signal included in the input mixed signal belongs to at each time. However, the performance of the GMM as a discriminator is low, and sufficient noise suppression performance cannot be obtained with the GMM.

  An example of an embodiment disclosed in the present application has been made in view of the above, and aims to improve noise suppression performance.

  In an example of the embodiment of the present application, an acoustic feature amount is extracted from an audio signal for learning. And an example of embodiment produces | generates the label information which matches the extracted acoustic feature-value and the mixing normal distribution of an audio | voice signal. In the exemplary embodiment, a normalized acoustic feature amount is extracted from a learning acoustic signal including a learning speech signal and a learning noise signal. In the exemplary embodiment, the speech model is learned using the generated label information and the extracted normalized acoustic feature amount.

  Further, in an example of the embodiment of the present application, the speech model learned by the speech model learning method is stored in the speech model storage unit. And an example of embodiment extracts an acoustic feature-value from the mixed acoustic signal containing an audio | voice signal and a noise signal. And an example of embodiment extracts the acoustic feature-value normalized from the mixed acoustic signal. Then, in an example of the embodiment, the speech posterior probability is calculated using the speech model and the extracted normalized acoustic feature amount. In the exemplary embodiment, the noise signal in the mixed acoustic signal is suppressed using the calculated voice posterior probability and the mixed normal distribution of the voice signal.

  According to an exemplary embodiment disclosed in the present application, for example, noise suppression performance can be improved.

FIG. 1 is a diagram illustrating an example of a configuration of a speech model learning apparatus. FIG. 2 is a flowchart illustrating an example of a processing procedure of the first acoustic feature extraction unit of the speech model learning device. FIG. 3 is a flowchart illustrating an example of a processing procedure of the second acoustic feature extraction unit of the speech model learning device. FIG. 4 is a diagram illustrating an example of the configuration of the noise suppression device. FIG. 5 is a diagram illustrating an example of a configuration of a parameter estimation unit of the noise suppression device. FIG. 6 is a flowchart illustrating an example of a processing procedure of the parameter estimation unit of the noise suppression device. FIG. 7 is a flowchart showing an example of a subroutine of the trust data selection process by the parameter estimation unit of the noise suppression device. FIG. 8 is a diagram illustrating an example of the configuration of the noise suppression unit of the noise suppression device. FIG. 9 is a flowchart illustrating an example of a processing procedure of the noise suppression filter estimation unit of the noise suppression device. FIG. 10 is a flowchart illustrating an example of a processing procedure of the noise suppression filter application unit of the noise suppression device. FIG. 11 is a diagram illustrating an example of the effect according to the embodiment. FIG. 12 is a diagram illustrating an example of a computer that realizes a speech model learning device and a noise suppression device by executing a program.

[Embodiment]
Hereinafter, embodiments of a speech model learning method, a noise suppression method, a speech model learning device, a noise suppression device, a speech model learning program, and a noise suppression program disclosed in the present application will be described. The following embodiments are merely examples, and do not limit the technology disclosed by the present application. Moreover, you may combine suitably embodiment shown below and other embodiment in the range with no contradiction.

In the following embodiments, for example, when “^ A” is written for A which is a vector or a scalar, it is equivalent to “a symbol with“ ^ ”immediately above“ A ””. When “￣A” is described, it is equivalent to “a symbol in which“ ￣ ”is written immediately above“ A ””. Also, when “A” is a vector, it is expressed as “vector A”, when “A” is a scalar, it is simply expressed as “A”, and when “A” is a set. And “set A”. For example, the function f of the vector A is expressed as f (vector A). Note that the matrix A ⁻¹ represents the inverse matrix of the matrix A with respect to the matrix A.

  In the following embodiments, a classifier based on a deep neural network (DNN) is introduced as a classifier for an audio signal. DNN is a kind of multi-layer perceptron, and a normal multi-layer perceptron has about three identification layers, whereas in the embodiment, it has more than three identification layers and constructs a deeper network. Specifically, each identification layer is learned by a restricted Boltzmann Machine (RBM), and then the RBM of each identification layer is connected to adjust the parameters of the entire network, thereby having a deep identification layer. A neural network can be constructed. By providing such a deep discrimination layer, the discrimination performance of the audio signal can be enhanced.

  In order to introduce a DNN voice signal discriminator to noise suppression, it is necessary to associate each node included in the DNN output layer with each element distribution of the GMM of the voice signal. For this purpose, first, a distribution label indicating which element distribution included in the GMM of the audio signal the audio signal without noise at each time belongs is generated. Thereafter, the DNN of the audio signal is learned using the mixed signal of the audio signal and the noise signal and the distribution label. By using such a method, it is possible to associate each element of the GMM of the audio signal with each node of the output layer of the DNN of the audio signal.

  Further, by using the DNN of the audio signal, the discrimination performance of the audio signal included in the input mixed signal is improved, the accuracy of extracting the audio signal and the noise signal from the input mixed signal, the speaker adaptation parameter, and the noise It is possible to improve the estimation accuracy with the suppression filter.

  Regarding DNN, reference 1 “A. Mohamed, G. Dahl, G. Hinton,“ Acoustic Modeling Using Deep Belief Networks. ”, IEEE Transactions on Audio, Speech, and Language Processing, vol. 20, no1., Pp. 14-22, 2012. ”, Reference 2“ Yotaro Kubo, “Pattern Recognition by Deep Learning”, Information Processing, vol. 54, no. 5, pp. 500-508, April 2013. ” .

(Configuration of speech model learning device)
FIG. 1 is a diagram illustrating an example of a configuration of a speech model learning apparatus. The speech model learning device 100 is connected to a speech GMM storage device 300 and a speech DNN storage device 400. The voice GMM storage device 300 stores a voice GMM 300a. The voice DNN storage device 400 stores a voice DNN 400a including a weight matrix W _j that is a parameter learned by a voice DNN learning unit 140 described later and a bias vector v _j . The speech model learning apparatus 100 receives a learning speech signal O ^clean _{τ, a} learning speech signal O ^clean _τ and a learning mixed signal O ^noisey _τ mixed with a learning noise signal, and a weight matrix W that is a DNN parameter. _j and the bias vector v _j are output. The speech model learning apparatus 100 includes a first acoustic feature extraction unit 110, a second acoustic feature extraction unit 120, a maximum likelihood distribution estimation unit 130, and a speech DNN learning unit 140.

The first acoustic feature extraction unit 110 receives the learning speech signal O ^clean _τ , and uses a learning logarithm that is a feature amount for obtaining the correspondence distribution label Lab _t used for learning the speech DNN from the learning speech signal O ^clean _τ. Extract mel spectrum vector O ^clean _t .

FIG. 2 is a flowchart illustrating an example of a processing procedure of the first acoustic feature extraction unit of the speech model learning device. The process of the first acoustic feature extraction unit 110 will be described with reference to FIG. First, the first acoustic feature extraction unit 110 moves the start point of the learning speech signal O ^clean _τ (τ is a sampling point of a discrete signal) in the time axis direction with a constant time width in the frame cutout process for a fixed time. A long acoustic signal is cut out as a frame (step S110a). For example, the first acoustic feature extraction unit 110 converts the acoustic signal O ^clean _{τ, n} of Frame = 400 sample points (16,000 Hz × 25 ms) into Shift = 160 sample points (16,000 Hz × 10 ms). Cut out while moving the start point. Here, t represents the frame number, and n represents the nth sample point in the frame. At that time, the first acoustic feature extraction unit 110, for example, cut out by multiplying a window function w _n, such as Hamming window shown in the following equation (1).

Thereafter, the first acoustic feature extraction unit 110 performs fast Fourier transform processing of M points (M is a power of 2 and a value equal to or greater than Frame, for example, M = 512) for the acoustic signal O ^clean _{t, n} . and a vector of complex spectrum _{^{Spc clean t = {Spc clean t}} , 0, ···, Spc clean t, m, ···, Spc clean t, M-1} to obtain a ^T (m is the number of frequency bins (Step S110b). Note that {·} ^T represents transposition of a matrix or a vector. Next, mel filter bank analysis processing (step S110c) and logarithmization processing (step S110d) are applied to the absolute value of each Spc ^clean _{t, m} , and an R-dimensional (for example, R = 24) log mel spectrum is obtained. Vector O ^clean _t = {O ^clean _{t, 0} ,..., O ^clean _{t, r} ,..., O ^clean _{t, R−1} } ^T is calculated (r is an element of the vector O ^clean _t number). As a result, the first acoustic feature extraction unit 110 outputs a vector O ^clean _t as a logarithmic mel spectrum for learning.

The second acoustic feature extraction unit 120 receives the learning mixed signal O ^noisey _{τ obtained} by mixing the learning speech signal O ^clean _τ and the learning noise signal, and performs speech model learning from the learning mixed signal O ^noisey _τ. The learning normalized log mel spectrum vector O ^noise _t is extracted.

FIG. 3 is a flowchart illustrating an example of a processing procedure of the second acoustic feature extraction unit of the speech model learning device. The process of the second acoustic feature extraction unit 120 will be described with reference to FIG. Second acoustic feature extraction unit 120 in step ^{S120a～S120d,} against ^{O noisy} _tau, shown in FIG. ^2, performing the same processing as each step S110a~S110d performed on ^{O clean} _tau.

Next, the second acoustic feature extraction unit 120 applies a normalization process to the logarithmic mel spectrum of the learning mixed signal O ^noisey _τ obtained by the logarithmic process of step S120d (step S120e). Specifically, the second acoustic feature extraction unit 120 uses the mean and standard deviation of the log mel spectrum of the learning mixed signal O ^noisy mixed signals for learning obtained from whole log mel spectrum of _tau O ^noisy _tau, learning use mixed signal ^{O noisy} averaged logarithmic Mel spectrum of _tau 0, normalized to unit variance.

Next, the second acoustic feature extraction unit 120 calculates the primary and secondary regression coefficients of the log mel spectrum of the learning mixed signal O ^noisey _τ normalized by the normalization process in step S120e, of the vector ^O of 3R dimensions together with logarithmic Mel spectrum of the learning mixed signal _{^{O noisy τ norm t = {O}} norm t, 0, ···, O norm t, r, ···, O norm t, _3R-1 } Regression coefficient provision processing that constitutes ^T is executed (step S120f). Thereafter, the second acoustic feature extraction unit 120, the front and rear Z frame {t-Z in step vector regression coefficients have been granted by the regression coefficient imparting treatment of S120f ^{O norm} _t frame t, · · ·, t, · · ·, t + Z} amount corresponding bound 3R × (2R + 1) dimensional vector ^{O norm} _t = {vector _{^{O norm t-Z T, ···}} , vector ^{O norm} _{t ^T,} ···, vector ^O _norm ^{t +} Z ^{T} T} Is executed (for example, Z = 5) (step S120g). Consequently, the second acoustic feature extraction unit 120 outputs the vector ^{O norm} _t of the learning normalized logarithmic Mel spectrum.

The maximum likelihood distribution estimation unit 130 uses the learning log mel spectrum vector O ^clean _t , which is the output of the first acoustic feature extraction unit 110, and the speech GMM 300a stored in the main memory of the speech GMM storage device 300. , The corresponding distribution label Lab _t is obtained.

The maximum likelihood distribution estimation unit 130 estimates the corresponding distribution label Lab _t used for learning of the speech DNN using the learning log mel spectrum vector O ^clean _t and the speech GMM 300a according to the following equation (2).

In the above equation (2), k is a normal distribution number included in the speech GMM 300a and takes the maximum value K. K is the total normal distribution number. For example, K = 512. In the above equation (2), w _{SI, k} is the mixing weight of the speech GMM 300a, the vector μSI _{, k} is an average vector of the speech GMM 300a, and the vector ΣSI _{, k} is a diagonal dispersion matrix of the speech GMM 300a. Each parameter, w _{SI, k} , vector μ _{SI, k} , and vector ΣSI _{, k,} is estimated in advance using speech data for learning of many speakers. In the above equation (2), the function N (•) is a probability density function of a multidimensional normal distribution given by the following equation (3). In the above equation (2), max {·} when k is scanned in the range of 1 ≦ k ≦ K is the corresponding distribution label Lab _t .

Voice DNN learning unit 140 uses the corresponding distribution labels Lab _t and vector ^{O noisy} _t of the learning normalized logarithmic Mel spectrum, learns the weight matrix _{W j} and the bias vector _{v j} is a parameter of the speech DNN400a. The speech DNN learning unit 140 uses the correspondence distribution label Lab _t estimated by the maximum likelihood distribution estimation unit 130 and the learning normalized log mel spectrum vector O ^noise _t calculated by the second acoustic feature extraction unit 120. Thus, a DNN having a hidden layer of J is learned as the speech DNN 400a (for example, J = 5). The general learning method of DNN is as shown in the above-mentioned literature 1 and literature 2.

The voice DNN learning unit 140 outputs the weight matrix W _j and the bias vector v _j that are parameters of the voice DNN 400a to the voice DNN storage device 400 and stores them in the main memory. The weight matrix W _j is a D _j × D _j−1 dimensional matrix, and the bias vector v _j is a D _j dimensional vertical vector (for example, D ₀ = 3R × (2Z + 1), D _j = 2048 ( j = 1,..., J-1), D _j = K).

(Configuration of noise suppression device)
FIG. 4 is a diagram illustrating an example of the configuration of the noise suppression device. The noise suppression device 200 is connected to a voice GMM storage device 300 and a voice DNN storage device 400. The noise suppression apparatus 200 receives an input mixed signal O _τ mixed with a speech signal and a noise signal, and outputs a noise suppression signal ^ S _τ that is estimated to be suppressed in the input mixed signal O _τ . The noise suppression apparatus 200 includes a first acoustic feature extraction unit 210, a second acoustic feature extraction unit 220, a parameter estimation unit 230, and a noise suppression unit 240.

The first acoustic feature extractor 210, the audio signal and the noise signal as input an input mixed signal O _tau mixing, a vector of complex spectrum is a feature amount for the practice of the noise suppression for the input mixed signal O _tau Spc A logarithmic mel spectrum vector O _t of _t and the input mixed signal O _τ is extracted. The first acoustic feature extraction unit 210 has the same processing function as the first acoustic feature extraction unit 110 of the speech model learning device 100.

The second acoustic feature extraction unit 220 receives the input mixed signal O _τ and receives a normalized log mel spectrum vector O ^DNN _t that is a feature amount for calculating the speech posterior probability P _{t, k} from the input mixed signal O _τ. To extract. The second acoustic feature extraction unit 220 has the same processing function as the second acoustic feature extraction unit 120 of the speech model learning device 100.

The parameter estimation unit 230 is extracted by the logarithmic mel spectrum vector O _t extracted by the first acoustic feature extraction unit 210, the speech GMM 300 a stored in the speech GMM storage device 300, and the second acoustic feature extraction unit 220. Using the normalized log mel spectrum vector O ^DNN _t and the speech DNN 400a stored in the speech DNN storage device 400, the speaker adaptation parameter vector b, and the noise GMM parameter set λ, which is a noise probability model, Is estimated.

Speech GMM 300a composed of parameters estimated from multi-speaker learning speech data is referred to as speaker independent (SI) GMM, and is composed of parameters estimated from specific speaker learning speech data. The voice GMM to be performed is called a speaker dependent (SD) GMM. Since it is not practical to learn the speaker independent GMM using the speech data for learning of a specific speaker, a speaker dependent GMM is obtained by speaker adaptation processing. That is, the average vector μ _{SD, k} of the speaker-dependent GMM is _obtained by converting the average vector μ _{SI, k} of the speaker independent GMM by speaker adaptation processing according to the following equation (4).

  In the above equation (4), the vector b is a speaker adaptation parameter and is an R-dimensional vector. The vector b is an independent parameter for the normal distribution number k included in the speech GMM 300a. On the other hand, the noise GMM is given by the following equation (5).

In the above equation (5), l is the number of the normal distribution included in the noise GMM, and L is the total number of normal distributions (for example, L = 4). Further, the vector N _t is a log mel spectrum of noise, and p _N (vector N _t ) is the likelihood of the noise GMM. Further, w _{N, l} is a noise GMM mixing weight, vector μ _{N, l} is an average vector of noise GMM, and matrix Σ _{N, l} is a diagonal dispersion matrix of noise GMM. Hereinafter, the parameter set of the noise GMM is defined as λ = {w _{N, l} , vector μ _{N, l} , matrix Σ _{N, l} }.

  The parameter estimation unit 230 estimates the speaker adaptation parameter vector b and the noise GMM parameter set λ using the EM algorithm. The EM algorithm is a method used for parameter estimation of a certain probability model, and maximizes the cost function with an Expectation-step (E-step) for calculating an expected value of a cost function (log likelihood function) of the probability model. The parameter is optimized by repeating Maximization-step (M-step) until the convergence condition is satisfied.

  Furthermore, a detailed configuration of the parameter estimation unit 230 illustrated in FIG. 4 will be described. FIG. 5 is a diagram illustrating an example of a configuration of a parameter estimation unit of the noise suppression device. As shown in FIG. 5, the parameter estimation unit 230 includes an initialization unit 231, a probability and signal estimation unit 232, a confidence data selection unit 233, a speaker adaptive parameter estimation unit 234, a noise GMM estimation unit 235, and a convergence determination unit 236. Have.

  FIG. 6 is a flowchart illustrating an example of a processing procedure of the parameter estimation unit of the noise suppression device. The process of the parameter estimation unit 230 will be described with reference to FIG. First, the initialization unit 231 initializes the repetition index of the EM algorithm as i = 1 (step S230a). Next, the initialization unit 231 executes initial value estimation processing for estimating the speaker adaptation parameter vector b in the EM algorithm and the initial values of the noise GMM parameter set λ using the following equations (6) to (11). (Step S230b). Here, U in the following equation (9) is the number of frames required for initial value estimation (for example, U = 10). Further, diag {·} in the following equation (9) represents that only the diagonal component of the matrix is calculated and the non-diagonal component is set to zero.

  In the above equation (9), the subscript i indicates a parameter in the i-th iterative estimation in the EM algorithm. Further, the vector 0 in the above equation (6) is an R-dimensional vertical vector whose element is 0. Further, GaussRand (•) in the above equation (10) is a function for generating normal random numbers.

Next, the probability and signal estimation unit 232 uses the normalized log mel spectrum vector O ^DNN _t , the weight matrix W _j and the bias vector v _j that are parameters stored in the speech ^DNN 400a, and the following equation (12): The speech posterior probability calculation process for calculating the speech posterior probability P _{t, k} is executed by the expression (15) (step S230c).

In the above equation (14), W _{j, k, k ′} are elements of the weight matrix W _j , v _{j, k} are elements of the bias vector v _j , and in the above equation (15), O ^DNN _{t , K} are elements of the vector O ^DNN _t .

Next, the probability and signal estimation unit 232 includes (i-1) a speaker adaptation parameter vector b ^(i-1) in the (i-1) th iteration estimation, and a noise GMM parameter set in the (i-1) th iteration estimation. Using λ ⁽ⁱ⁻¹⁾ and the parameters of the speech GMM 300a, a mixed signal GMM generation process constituting the GMM of the logarithmic mel spectrum vector O _t as shown in the following equation (16) is executed (step S230d). .

In the above equation (16), p _o ⁽ⁱ⁾ (vector O ^t ) is the likelihood of the log mel spectrum vector O _t generated in the mixed signal GMM generation processing in step S230d with respect to the speech GMM 300a. . Also, w _{O, k, l} ⁽ⁱ⁾ , vector μ _{O, k, l} ⁽ⁱ⁾ , matrix Σ _{O, k, l} ⁽ⁱ⁾ are respectively speaker adaptations in the (i−1) -th iteration estimation. GMM mixture weight, average vector, pair of logarithmic mel spectrum vector Ot generated from parameter set vector b ^(i-1) , noise GMM parameter set λ ^(i-1), and parameters of speech GMM 300a It is an angular dispersion matrix and is given by the following equations (17) to (20).

In the above equation (18), the logarithmic function log (•) and the exponential function exp (•) are calculated for each element of the vector. In the above equations (18) and (20), the vector 1 is an R-dimensional vertical vector in which all elements are 1. In the above equation (19), H _{k, l} ⁽ⁱ⁾ is a Jacobian matrix of the function h (•).

Next, the probability and signal estimation unit 232 calculates an expected value by calculating the expected value of the cost function Qo (•) of the probability model of the logarithmic mel spectrum vector O _{t in} the i-th iterative estimation using the following equation (21). The process is executed (EM algorithm E-step) (step S230e).

In the above (21), the vector _{O 0: T-1 = {} O 0, ···, O t, ··· O T-1} is. In the above equation (21), T is the total number of frames of the logarithmic mel spectrum vector O _t . Further, in the above equation (21), P _{t, k, l} ⁽ⁱ⁾ is changed to the normal distribution number k of the speech GMM 300a and the normal distribution number l of the noise GMM in the frame t by the following equations (22) and (23). This is the voice posterior probability given to the user.

  Note that M-step of the EM algorithm corresponds to the signal estimation process in step S230f, the confidence data selection process in step S230g, the speaker adaptive parameter estimation process in step S230h, and the noise GMM parameter estimation process in step S230i.

In step S230f, the probability and signal estimation unit 232 uses the logarithmic mel spectrum vector S _{t of the} clean speech used to update the speaker adaptation parameter vector b ⁽ⁱ⁾ and the noise GMM parameter set λ ^(i). and ^(i), and a noise of the logarithmic Mel spectrum vector _N ^{t (i),} estimated from the logarithmic Mel spectrum vector _{O t.} The logarithmic mel spectrum vector S _t ^{(i) of} clean speech and the noise mel spectrum vector N _t ⁽ⁱ⁾ of noise are estimated by the following equations (24) and (25).

Next, the trust data selection unit 233 uses the estimated log mel spectrum vector {circumflex over ^(S) } of clean speech used when estimating the speaker adaptation parameter vector b ⁽ⁱ⁾ and the noise GMM parameter set λ ^(i). A trust data selection process for selecting _t ⁽ⁱ⁾ and an estimated log mel spectrum of noise ^ N _t ⁽ⁱ⁾ is executed (step S230g).

FIG. 7 is a flowchart showing an example of a subroutine of the trust data selection process by the parameter estimation unit of the noise suppression device. In the reliability data selection process, based on the result of determining which of clean speech and noise is dominant in all frames, if clean speech is dominant, each frame number t is set to a set T of clean speech signal frames. If it is stored in _S ⁽ⁱ⁾ and noise is dominant, each frame number t is stored in a set T _N ⁽ⁱ⁾ of noise frames. As shown in FIG. 7, first, the reliable data selection unit 233 calculates SNR _t ⁽ⁱ⁾ that is an SN ratio in each frame t by the following equation (26).

In the above (26), ^ S t, r (i) is an element of the vector of the estimated logarithmic Mel spectrum of the clean speech ^ S t (i) in the frame t, ^ N t, r ( i) is This is an element of the vector ^ N t (i) of the estimated log mel spectrum of noise in frame t. Then, the reliability data selection unit 233 applies k-mean clustering to SNR t (i) , which is the S / N ratio in each frame t, obtained by the above equation (26), so that the SNR t ( i) is classified into two classes C = 0, 1, and the average SN ratio of each class is defined as AveSNR c (i) (step S230g-1).

Then, the reliability data selection unit 233 determines whether or not AveSNR _{c = 0} ⁽ⁱ⁾ ≧ AveSNR _{c = 1} ⁽ⁱ⁾ in each frame t (step S230g-2). When the trust data selection unit 233 determines that AveSNR _{c = 0} ⁽ⁱ⁾ ≧ AveSNR _{c = 1} ⁽ⁱ⁾ in the frame t, the process moves to step S230g-3. On the other hand, if the trust data selection unit 233 determines that AveSNR _{c = 0} ⁽ⁱ⁾ <AveSNR _{c = 1} ⁽ⁱ⁾ in the frame t, the process moves to step S230g-6.

In step S230g-3, the trust data selection unit 233 determines that SNR _t ⁽ⁱ⁾ in each frame t is SNR _t ⁽ⁱ⁾ ∈ {C = 0}, that is, SNR _t ⁽ⁱ⁾ is a set {C = 0} ( Whether or not it belongs to the cluster (C = 0). The trust data selection unit 233 shifts the processing to step S230g-4 for the frame t determined to be SNR _t ⁽ⁱ⁾ ε {C = 0}. On the other hand, the trust data selection unit 233 moves the process to step S230g-5 for the frame t determined to be SNR _t ⁽ⁱ⁾ ε {C = 1}.

In step S230g-4, the reliable data selection unit 233 stores the frame number t determined in step S230g-3 in the set T _S ^{(i) of} clean speech signal frames. On the other hand, in step S230g-5, the reliable data selection unit 233 stores the frame number t determined in step S230g-3 in the noise signal frame set T _N ⁽ⁱ⁾ .

On the other hand, in step S230g-6, the trust data selection unit 233 determines that SNR _t ⁽ⁱ⁾ in each frame t is SNR _t ⁽ⁱ⁾ ∈ {C = 1}, that is, SNR _t ⁽ⁱ⁾ is a set {C = 1. } (C = 1 cluster). The trust data selection unit 233 moves the process to step S230g-7 for the frame t determined to be SNR _t ⁽ⁱ⁾ ε {C = 1}. On the other hand, the trust data selection unit 233 moves the process to step S230g-8 for the frame t determined to be SNR _t ⁽ⁱ⁾ ε {C = 0}.

In step S230g-7, the reliable data selection unit 233 stores the frame number t determined in step S230g-6 in the clean speech signal frame set T _S ⁽ⁱ⁾ . On the other hand, in step S230g-8, the reliable data selection unit 233 stores the frame number t determined in step S230g-6 in the noise signal frame set T _N ⁽ⁱ⁾ . When the processes of steps S230g-4, S230g-5, S230g-7, and S230g-8 are completed, the reliable data selection unit 233 returns the process to the process of the parameter estimation unit 230 of the noise suppression device illustrated in FIG.

Next, the speaker adaptive parameter estimation unit 234 uses the speech posterior probability P _{t, k} obtained in the speech posterior probability calculation process in step S230c and the log mel spectrum ^ of the clean speech estimated in the signal estimation process in step S230f. Using a set T _S ⁽ⁱ⁾ of clean speech signal frames estimated in S _t ⁽ⁱ⁾ and the reliability data selection process in step S230g, a speaker adaptation parameter vector b ^{(i) according to the following} equation (27 ^): The speaker adaptation parameter estimation processing for updating the is executed (step S230h).

Next, the noise GMM estimator 235 calculates the speech posterior probability P _{t, l} ⁽ⁱ⁾ obtained in the expected value calculation process in step S230e and the log mel spectrum vector of the noise estimated in the signal estimation process in step S230f ^ and _N ^{t (i),} using a set _T ^{N (i)} of the noise signal frame estimated by reliable data selection processing in step S230g, the following (28) to (30), a parameter set of the noise GMM A noise GMM parameter estimation process for updating λ ⁽ⁱ⁾ is executed (step S230i).

Next, the convergence determination unit 236 executes a convergence determination process for determining whether or not a predetermined convergence condition is satisfied (step S230j). If the predetermined convergence condition is satisfied, the convergence determination unit 236 sets the vector b = b (i) and ends the processing of the parameter estimation unit 230. On the other hand, when the predetermined convergence condition is not satisfied, the convergence determination unit 236 increments i by 1 (i ← i + 1) (step S230k), and moves the process to step S230d. The predetermined convergence condition is expressed by the following equation (31). In the following equation (31), Q _O (•) is defined by the above equation (21). In the following equation (31), η = 0.0001.

Further, the detailed configuration of the noise suppression unit 240 shown in FIG. 4 will be described. FIG. 8 is a diagram illustrating an example of the configuration of the noise suppression unit of the noise suppression device. The noise suppression unit 240 includes a complex spectrum vector Spc _t , a log mel spectrum vector O _t , a speech GMM 300 a, a speaker adaptation parameter vector b, a noise GMM parameter set λ, and a speech posterior probability P _t, configure the noise suppression filter using the _k, obtain noise suppression signal ^ S _tau to suppress noise.

As illustrated in FIG. 8, the noise suppression unit 240 includes a noise suppression filter estimation unit 241 and a noise suppression filter application unit 242. The noise suppression filter estimation unit 241 receives a logarithmic mel spectrum vector Ot, a speech GMM 300a, a speaker adaptation parameter vector b, a noise GMM parameter set λ, and a speech posterior probability P _{t, k.} The suppression filter F _{t, m} ^Lin is estimated. Noise suppression filter applying unit 242 obtains a vector Spc _t complex spectrum, the noise suppression filter _F ^t, as input and ^{m Lin,} the noise suppression signal ^ S _tau to suppress noise.

  FIG. 9 is a flowchart illustrating an example of a processing procedure of the noise suppression filter estimation unit of the noise suppression device. First, the noise suppression filter estimation unit 241 determines the GMM parameters of the logarithmic mel spectrum vector Ot from the speech GMM 300a, the speaker adaptation parameter vector b, and the noise GMM parameter set λ by the following equation (32): Probability model generation processing is generated as shown in equation (35) (step S241a).

Next, the noise suppression filter estimation unit 241 calculates the posterior probabilities P _{t, k, l} , the GMM parameters of the log mel spectrum vector O _t , and the log mel spectrum by the following equations (36) and (37). Probability calculation processing is performed using the vector O _t and the speech posterior probability P _{t, k} (step S241b).

Next, the noise suppression filter estimation unit 241 includes an average vector μ _{SD, k of the} speaker-dependent (SD) GMM generated from the average vector μ _{SI, k} of the speech GMM 300a and the vector b of the speaker adaptation parameter, The noise suppression filter F _{t, r} ^Mel on the mel frequency axis is _{expressed as} follows using the average vector μ _{N, l of the} noise GMM included in the parameter set λ of the noise GMM and the posterior probability P _{t, k, l.} A noise suppression filter estimation process for estimation as shown in equation (38) is executed (step S241c). The following equation (38) is a notation for each vector element.

Next, the noise suppression filter estimation unit 241 performs a noise suppression filter conversion process for converting the noise suppression filter F _{t, r} ^Mel on the mel frequency axis to the noise suppression filter F _{t, r} ^Lin on the linear frequency axis. It executes (step S241d). The process of converting the noise suppression filter F _{t, r} ^Mel on the mel frequency axis to the noise suppression filter F _{t, r} ^Lin on the linear frequency axis is obtained by applying cubic spline interpolation to the mel frequency axis. The value of the noise suppression filter on the linear frequency axis is estimated. When step S241d ends, the processing of the noise suppression filter estimation unit 241 ends.

FIG. 10 is a flowchart illustrating an example of a processing procedure of the noise suppression filter application unit of the noise suppression device. First, the noise suppression filter applying unit 242, the noise suppression filter _F ^t For complex spectrum vector Spc ^_t, the ^{m Lin,} by multiplying as follows (39) equation, the noise-suppressed complex spectrum ^ S _A filtering process for obtaining _{t and m} is executed (step S242a). The following equation (39) is a notation for each vector element.

Next, the noise suppression filter application unit 242 performs inverse fast Fourier change processing for obtaining the noise-suppressed speech ^ S _{t, n} in the frame t by applying inverse fast Fourier transform to the complex spectrum ^ S _{t, m} . It executes (step S242b). Next, the noise suppression filter applying unit 242, the noise reduced speech _{^ S t} of each frame _t, a _n, as follows (40) and (41) below, in conjunction with releasing the window function _{w n} Then, the waveform concatenation process for obtaining the continuous noise-suppressed speech ^ s _τ is executed (step S242c). When step S242c ends, the processing of the noise suppression filter application unit 242 ends.

[Effects of the embodiment]
In order to show the effect of the embodiment, an example is shown in which an acoustic signal in which a voice signal and a noise signal are mixed is input to the noise suppression apparatus 200 of the embodiment and noise suppression is performed. Hereinafter, experimental methods and results will be described.

  In the experiment, evaluation was performed using a speech recognition database under a noisy environment called AURORA4. The AURORA4 evaluation data set consists of A: voice without noise, B: voice mixed with six kinds of noise, C: voice without noise recorded with different microphones, and D: six kinds of noise recorded with different microphones. Is composed of 4 sets of mixed audio. For details of AURORA4, see Reference 3 “N. Parihar, J. Picone, D. Pearce, HG Hirsch,“ Performance analysis of the Aurora large vocabulary baseline system. ”In Proceedings of the European Signal Processing Conference, Vienna, Austria, 2004. . ".

  The audio data of AURORA 4 is a monaural signal that is discretely sampled at a sampling frequency of 16,000 Hz and a quantization bit number of 16 bits. For the acoustic signal based on this audio data, the time length of one frame is 25 ms (Frame = 400 sample points), and the start point of the frame is moved every 10 ms (Shift = 160 sample points) to perform acoustic feature extraction. .

As the speech GMM 300a, a GMM with a mixed distribution number K = 512 having an R = 24-dimensional logarithmic mel spectrum as an acoustic feature quantity was used, and learning was performed using speech data for learning with no AURORA4 noise mixing. L = 4 was given to the number of mixed distributions of the noise GMM. The audio DNN 400a includes a total D ₀ = 3R × (2Z + 1) including R = 24-dimensional logarithmic mel spectrum, its primary and secondary regression coefficients, and feature quantities of Z = 5 frames before and after the current frame. ) = 792 dimensional vectors as acoustic features, J = 5 hidden layers, D ₀ = 792 nodes in the input layer, D _j = 2048 (j = 1,..., 4) nodes in the hidden layer Then, using a DNN having D ₅ = K = 512 nodes in the output layer, learning was performed using learning speech data mixed with AURORA4 noise.

  Speech recognition was performed by a recognizer based on a finite state transducer. Details of the recognizer based on the finite state transducer are described in Reference 4 “T. Hori, et al.,“ Efficient WFST-based one-pass decoding with on-the-fly hypothesis rescoring in extremely large vocabulary continuous speech recognition. ”IEEE On ASLP, vol. 15, no. 4, pp. 1352-1365, May 2007. ”.

  The acoustic model uses DNN and has seven hidden layers. The number of nodes in each hidden layer is 2048. The number of nodes in the output layer is 3042. The acoustic feature amount for speech recognition is a 24-dimensional log mel spectrum analyzed by moving the start point of a frame every 10 ms (Shift = 1600 sample points) with a time length of one frame being 25 ms (Frame = 400) and its 1 This is a 792-dimensional vector in total including second and second order regression coefficients and feature quantities of the previous and next 5 frames around the current frame. The language model uses Tri-gram and the vocabulary is 5,000 words. The evaluation scale for speech recognition was the word error rate (WER) of the following equation (42). In the following equation (42), N is the total number of words, D is the number of dropped error words, S is the number of replacement error words, I is the number of insertion error words, and the smaller the WER value, the higher the speech recognition performance. .

  FIG. 11 is a diagram illustrating an example of the effect according to the embodiment. The “prior art” shown in FIG. 11 is described in Reference 5 “M. Fujimoto and T. Nakatani,“ A reliable data selection for model-based noise suppression using unsupervised joint speaker adaptation and noise model estimation. ”In Proceedings of ICSPCC '12 , pp. 4713-4716, Aug 2012. ”shows the result of noise suppression by the method disclosed in“ FIG. 11 shows a comparison of evaluation results of speech recognition of “no noise suppression”, “prior art”, and “embodiment”. As shown in FIG. 11, it can be seen that the embodiment can obtain higher noise suppression performance because the WER is small in the evaluation sets B and D including noise as compared with the related art.

  That is, according to the embodiment, in an environment where various types of noise exist, even if the noise signal included in the acoustic signal is non-stationary noise that follows a multimodal distribution, the noise signal is suppressed from the input acoustic signal. Thus, the target audio signal can be extracted with high quality.

[Other Embodiments]
In other embodiments, the frame cutout process of step S120a of steps S110a and 3 of Figure 2, as a window function w _n, besides Hamming window, utilizing a rectangular window, Hanning window, a window function, such as Blackman windows May be. In other embodiments, instead of the speech GMM 300a, another probability model such as a Hidden Markov Model (HMM) may be used as the probability model of the speech signal. In other embodiments, instead of the noise GMM, another probability model such as an HMM may be used as the noise signal probability model.

  In other embodiments, the speaker adaptation parameter vector b may be a parameter depending on the number k of the normal distribution included in the speech GMM 300a, as shown in the following equation (43).

In other embodiments, the trust data selection process shown in step S230g of FIG. 6 and FIG. 7 is executed using a predetermined threshold Th _SNR as shown in the following equation (44) instead of k-means clustering. May be.

Also, in other embodiments, the noise suppression filter estimation process in step S241c of FIG. 9, the (38) each posterior probability P _t as _{formula, k,} rather than the weighted average of _l, the maximum of the largest weight, that An estimation result weighted by the posterior probability P _{t, k, l} may be used. In this case, it is desirable that the maximum posterior probability P _{t, k, l} is sufficiently larger than the other posterior probabilities P _{t, k, l} .

(About device configuration of speech model learning device and noise suppression device)
Each component of the speech model learning apparatus 100 illustrated in FIG. 1 and the noise suppression apparatus 200 illustrated in FIG. 4 is functionally conceptual, and does not necessarily need to be physically configured as illustrated. That is, the specific forms of the distribution and integration of the functions of the speech model learning device 100 and the noise suppression device 200 are not limited to those shown in the figure, and all or a part of them can be changed into arbitrary units according to various loads and usage conditions. And can be configured to be functionally or physically distributed or integrated. For example, the speech model learning device 100 and the noise suppression device 200 may be an integrated device.

  In the embodiment, the speech model learning device 100 and the noise suppression device 200 are separate devices, and the first acoustic feature extraction unit 110 and the second acoustic feature extraction unit 120 of the speech model learning device 100 and the first of the noise suppression device 200 are used. The first acoustic feature extraction unit 210 and the second acoustic feature extraction unit 220 are different functional components. However, the present invention is not limited to this, and the first acoustic feature extraction unit 110 and the first acoustic feature extraction unit 210 and / or the second acoustic feature extraction unit 120 and the second acoustic feature extraction unit 220 are the same functional components. There may be.

  In the embodiment, the speech GMM storage device 300 and the speech DNN storage device 400 are separate from the speech model learning device 100 and the noise suppression device 200. However, the present invention is not limited to this, and the speech GMM storage device 300 and / or the speech DNN storage device 400 may be an apparatus integrated with the speech model learning device 100 and / or the noise suppression device 200.

  Each processing performed in the speech model learning device 100 and the noise suppression device 200 is realized in whole or in part by a processing device such as a CPU (Central Processing Unit) and a program that is analyzed and executed by the processing device. May be. Each process performed in the speech model learning device 100 and the noise suppression device 200 may be realized as hardware by wired logic.

  In addition, among the processes described in the embodiment, all or a part of the processes described as being automatically performed can be manually performed. Alternatively, all or some of the processes described as being manually performed among the processes described in the embodiments can be automatically performed by a known method. In addition, the above-described and illustrated processing procedures, control procedures, specific names, and information including various data and parameters can be changed as appropriate unless otherwise specified.

(About the program)
FIG. 12 is a diagram illustrating an example of a computer that realizes a speech model learning device and a noise suppression device by executing a program. The computer 1000 includes a memory 1010 and a CPU 1020, for example. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. In the computer 1000, these units are connected by a bus 1080.

  The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012. The ROM 1011 stores a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1031. The disk drive interface 1040 is connected to the disk drive 1041. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1041. The serial port interface 1050 is connected to a mouse 1051 and a keyboard 1052, for example. The video adapter 1060 is connected to the display 1061, for example.

  The hard disk drive 1031 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, a program that defines each process of the speech model learning device 100 and the noise suppression device 200 is stored in, for example, the hard disk drive 1031 as a program module 1093 in which commands executed by the computer 1000 are described. For example, a program module 1093 for executing information processing similar to the functional configuration in the speech model learning device 100 and the noise suppression device 200 is stored in the hard disk drive 1031.

  The setting data used in the processing of the embodiment is stored as program data 1094 in, for example, the memory 1010 or the hard disk drive 1031. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1031 to the RAM 1012 as necessary, and executes them.

  Note that the program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1031, but may be stored in, for example, a removable storage medium and read out by the CPU 1020 via the disk drive 1041 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and the program data 1094 may be read by the CPU 1020 via the network interface 1070.

  The above-described embodiments and other embodiments are included in the invention disclosed in the claims and equivalents thereof as well as included in the technology disclosed in the present application.

100 speech model learning device 110 first acoustic feature extraction unit 120 second acoustic feature extraction unit 130 maximum likelihood distribution estimation unit 140 speech DNN learning unit 200 noise suppression device 210 first acoustic feature extraction unit 220 second acoustic feature extraction unit 230 Estimation unit 231 Initialization unit 232 Probability and signal estimation unit 233 Reliability data selection unit 234 Speaker adaptive parameter estimation unit 235 Noise GMM estimation unit 236 Convergence determination unit 240 Noise suppression unit 241 Noise suppression filter estimation unit 242 Noise suppression filter application unit 300 Voice GMM storage device 300a Voice GMM
400 voice DNN storage device 400a voice DNN
1000 Computer 1010 Memory 1020 CPU

Claims (12)

A speech model learning method executed by a speech model learning device,
A feature extraction step for learning that extracts an acoustic feature from a speech signal for learning;
An audio label generation step of generating label information that associates the acoustic feature amount extracted by the learning feature amount extraction step with a mixed normal distribution of the audio signal;
A learning normalized feature amount extraction step of extracting a normalized acoustic feature amount from a learning acoustic signal including the learning speech signal and the learning noise signal;
A speech model learning step of learning a speech model using the label information generated by the speech label generation step and the normalized acoustic feature amount extracted by the learning normalization feature amount extraction step. A speech model learning method characterized by The speech model learning step includes a mixed normal distribution of the speech signal and each node of the output layer of the deep neural network corresponding to the normalized acoustic feature amount extracted by the learning normalized feature amount extraction step. The speech model learning method according to claim 1, wherein the speech model is learned by associating the speech model. A noise suppression method performed by a noise suppression device,
A speech model storage step of storing the speech model learned by the speech model learning method according to claim 1 or 2 in a speech model storage unit;
A feature extraction step of extracting an acoustic feature amount from a mixed acoustic signal including an audio signal and a noise signal;
A normalized feature extraction step of extracting a normalized acoustic feature from the mixed acoustic signal;
A speech posterior probability calculation step of calculating a speech posterior probability using the speech model and the normalized acoustic feature amount extracted by the normalized feature amount extraction step;
A noise suppression step of suppressing the noise signal in the mixed acoustic signal using the speech posterior probability calculated by the speech posterior probability calculation step and a mixed normal distribution of the voice signal. Repression method. A signal estimation step of estimating the audio signal and the noise signal included in the mixed acoustic signal;
Speaker adaptation for estimating speaker adaptation parameters for adapting a mixed normal distribution of the speech signal to a speech speaker corresponding to the speech signal from the speech signal and the noise signal estimated by the signal estimation step. A parameter estimation step;
A noise mixed normal distribution generating step for generating a mixed normal distribution of noise signals from the noise signal estimated by the signal estimating step;
A mixed normal distribution generating step of generating a mixed normal distribution of the mixed acoustic signal from the mixed normal distribution of the speaker adaptation parameter and the voice signal and the mixed normal distribution of the noise signal;
An expected value calculation step of calculating an expected value of the speech signal and an expected value of the noise signal included in the mixed acoustic signal from the speech posterior probability and a mixed normal distribution of the mixed acoustic signal;
The signal estimation step, the speaker adaptation parameter estimation step, the noise mixed normal distribution generation step, the mixed normal distribution generation step, and the expected value calculation step are the expected values of the speech signal calculated by the expected value calculation step. 4. The noise suppression method according to claim 3, wherein processing is recursively repeated for the expected value of the speech signal and the expected value of the noise signal until the expected value of the noise signal satisfies a predetermined condition. A selection step of selecting a signal satisfying a predetermined condition from the voice signal and the noise signal estimated by the signal estimation step,
The speaker adaptation parameter estimation step estimates the speaker adaptation parameter from the voice signal and the noise signal selected by the selection step,
The noise suppression method according to claim 4, wherein the noise mixed normal distribution generation step generates a mixed normal distribution of the noise signal from the noise signal selected by the selection step. A feature extraction unit for learning that extracts an acoustic feature from a speech signal for learning;
An audio label generation unit that generates label information that associates the acoustic feature amount extracted by the learning feature amount extraction unit with the mixed normal distribution of the audio signal;
A learning normalized feature quantity extraction unit that extracts a normalized acoustic feature quantity from a learning acoustic signal including the learning speech signal and the learning noise signal;
A speech model learning unit that learns a speech model using the label information generated by the speech label generation unit and the normalized acoustic feature amount extracted by the learning normalized feature amount extraction unit. A speech model learning apparatus characterized by that. The speech model learning unit includes a mixed normal distribution of the speech signal and each node of the output layer of the deep neural network corresponding to the normalized acoustic feature amount extracted by the learning normalized feature amount extraction unit. The speech model learning apparatus according to claim 6, wherein the speech model is learned by associating the speech model. A speech model storage unit that stores the speech model learned by the speech model learning device according to claim 6;
A feature extraction unit that extracts an acoustic feature amount from a mixed acoustic signal including an audio signal and a noise signal;
A normalized feature quantity extraction unit for extracting a normalized acoustic feature quantity from the mixed acoustic signal;
A speech posterior probability calculation unit that calculates a speech posterior probability using the speech model and the normalized acoustic feature amount extracted by the normalized feature amount extraction unit;
A noise suppression unit comprising: a noise suppression unit that suppresses the noise signal in the mixed acoustic signal using the speech posterior probability calculated by the speech posterior probability calculation unit and a mixed normal distribution of the speech signal. apparatus. A signal estimation unit for estimating the audio signal and the noise signal included in the mixed acoustic signal;
Speaker adaptation for estimating speaker adaptation parameters for adapting a mixed normal distribution of the speech signal to speech speakers corresponding to the speech signal from the speech signal and the noise signal estimated by the signal estimation unit A parameter estimator;
From the noise signal estimated by the signal estimation unit, a noise mixed normal distribution generation unit that generates a mixed normal distribution of noise signals,
A mixed normal distribution generating unit that generates a mixed normal distribution of the mixed acoustic signal from the mixed normal distribution of the speaker adaptation parameter and the voice signal and the mixed normal distribution of the noise signal;
An expected value calculation unit for calculating an expected value of the speech signal and an expected value of the noise signal included in the mixed acoustic signal from the speech posterior probability and a mixed normal distribution of the mixed acoustic signal; and
The signal estimation unit, the speaker adaptive parameter estimation unit, the noise mixed normal distribution generation unit, the mixed normal distribution generation unit, and the expected value calculation unit are expected values of the speech signal calculated by the expected value calculation unit. The noise suppression device according to claim 8, wherein the processing is recursively repeated for the expected value of the speech signal and the expected value of the noise signal until the expected value of the noise signal satisfies a predetermined condition. A selection unit that selects a signal that satisfies a predetermined condition from the voice signal and the noise signal estimated by the signal estimation unit;
The speaker adaptation parameter estimation unit estimates the speaker adaptation parameter from the voice signal and the noise signal selected by the selection unit,
The noise suppression apparatus according to claim 9, wherein the noise mixed normal distribution generation unit generates a mixed normal distribution of the noise signal from the noise signal selected by the selection unit.   A speech model learning program for causing a computer to function as the speech model learning device according to claim 6.   A noise suppression program for causing a computer to function as the noise suppression device according to claim 8, 9 or 10.

Images