JP6521886B2 - Signal analysis apparatus, method, and program - Google Patents

Description

  The present invention relates to a signal analysis apparatus, method, and program, and more particularly to a signal analysis apparatus, method, and program for estimating parameters.

  The present invention addresses the problem of suppressing noise from speech signals. The noise mixed in the voice signal not only degrades the quality of voice communication but also causes the performance degradation of various voice processing such as voice recognition and voice conversion. Various speech enhancement methods have been proposed so far to solve this problem.

  Speech enhancement methods are roughly divided into unsupervised approach, supervised approach and semi-supervised approach.

  The supervised approach assumes the situation where the target speech and the target noise sample are obtained in advance, the semi-supervised approach takes the situation where only the target speech sample is obtained in advance, and the unsupervised approach does not obtain either. It is an emphasizing method. Further, the target to be emphasized is roughly classified into the case of a signal (or spectrum) and the case of a feature quantity. Typical examples of the supervised feature enhancement approach include the Vector Taylor Series (VTS) method, the Stereo Piecewise Linear Compensation for Environment (SPLICE), and the method using the Denoising Autoencoder (DAE).

  The VTS method is a method of constructing a conversion function from a noisy speech feature to a clean speech feature by first approximating a linear superposition process of speech and noise in a feature space.

  SPLICE models simultaneous probability density functions of features of noisy speech and clean speech with Gaussian mixture model (GMM), and cleans speech from noisy speech with GMM parameters learned using training samples This is a method of constructing a conversion function to feature quantities.

  The DAE method is a method of forming a conversion function between input and output by a deep layer neural network in which a noisy speech feature is input and a clean speech feature is output. These supervised speech enhancement approaches are quite powerful in known noise environments because they are based on discrimination models and criteria, but are not always effective in unknown noise environments. However, many methods have been proposed to compensate for the mismatch when speech or noise of learning data is different from that at the time of test.

  On the other hand, the method based on the semi-supervised non-negative matrix factorization (SSNMF) (non-patent document 1), which is a representative example of the semi-supervised signal enhancement approach, is under unknown noise environment It has been attracting attention in recent years as a powerful speech enhancement method. This method is based on the principle that it is possible to estimate the power spectrum of speech and noise by fitting the observation spectrum of each time instant with the non-negative coupling of the basis spectrum of noise pre-learned and the basis spectrum of noise.

P. Smaragdis, B. Raj, and M. Shashanka, “Supervised and semi-supervised separation of sounds from single-channel mixtures,” in Proc. Independent Component Analysis and Signal Separation, pp. 414-421, 2007.

  However, if the noise spectrum can be explained by the basis spectrum of speech or vice versa, the estimated spectrum may not correspond to the actual speech spectrum. For this reason, in order to eliminate the ambiguity of the decomposition of the speech spectrum and the noise spectrum, stronger restrictions that the speech spectrum must satisfy are required. Further, in the SSNMF method, even if the signal can be enhanced, there is no guarantee that the feature amount can be enhanced, so the enhancement processing is not always directly linked to the performance improvement of voice processing such as voice recognition and voice conversion based on the voice feature amount.

  The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a signal analysis device, method, and program capable of suppressing noise, emphasizing an audio signal and emphasizing a cepstrum feature. To aim.

  In order to achieve the above object, a signal analysis apparatus according to the present invention receives, as input, time-series data of an observation signal in which an audio signal and a noise signal are mixed, and displays an observation time frequency component of each time and each frequency It is defined by a time-frequency expansion unit outputting a spectrogram, the observation spectrogram output by the time-frequency expansion unit, a base spectrum representing a power spectrum at each base and each frequency of a speech signal learned in advance, and a cepstrum space Based on a parameter representing the probability distribution of the cepstral feature of the speech signal learned in advance, observation time frequency components of each time and each frequency, the base spectrum of the speech signal, and the power at each time of the speech signal Time period of each time and each frequency obtained from the activation parameter The speech based on a number component, a distance between the base spectrum of the noise signal and the sum of time-frequency components of each time and each frequency determined from the activation parameter, and the probability distribution of cepstrum feature of the speech signal The speech signal is reduced so as to reduce a criterion expressed using a regularization term that represents the likelihood of cepstral features of time frequency components at each time and frequency determined from the base spectrum of the signal and the activation parameter. And a parameter estimation unit configured to estimate the activation spectrum of the noise signal and the activation parameter of the noise signal.

  A signal analysis method according to the present invention is a signal analysis method in a signal analysis apparatus including a time-frequency expansion unit and a parameter estimation unit, wherein the time-frequency expansion unit is an observation in which an audio signal and a noise signal are mixed. An observation spectrogram representing observation time-frequency components of each time and each frequency is output with the time series data of the signal as input, and the parameter estimation unit learns in advance the observation spectrogram output by the time frequency expansion unit. At each time and each frequency, based on the basis spectrum representing the power spectrum at each base and each frequency of the speech signal, and the parameter representing the probability distribution of the cepstral feature of the pre-learned speech signal defined in cepstral space. An observation time frequency component, a base spectrum of the audio signal, and each time of the audio signal And the time frequency component of each time and frequency determined from the activation parameter representing the power, and the distance between the base spectrum of the noise signal and the sum of time frequency components of each time and frequency determined from the activation parameter And, based on the probability distribution of the cepstral feature of the audio signal, regularization representing the likelihood of the cepstral feature of the time-frequency component of each time and each frequency obtained from the base spectrum of the audio signal and the activation parameter The activation parameters of the speech signal, the basis spectrum of the noise signal and the activation parameters are estimated so as to reduce the criteria expressed using terms.

  Further, a program of the present invention is a program for causing a computer to function as each unit constituting the above-described signal analysis device.

  As described above, according to the signal analysis apparatus, method, and program of the present invention, the observation time frequency components of each time and each frequency, the base spectrum of the voice signal, and the power at each time of the voice signal A time-frequency component of each time and frequency determined from an activation parameter representing the distance, and a distance from a sum of time-frequency components of each time and frequency determined from the base spectrum of the noise signal and the activation parameter; Based on the probability distribution of the cepstral feature of the audio signal, using a regularization term representing the likelihood of the cepstral feature of the time-frequency component of each time and frequency determined from the base spectrum of the audio signal and the activation parameter To reduce the criteria expressed by It said activation parameter of the signal, by estimating the said base spectrum and the activation parameters of the noise signal, to suppress noise, as well as emphasizing the speech signal, can be emphasized cepstrum characteristic quantity.

It is a block diagram showing functional composition of a signal analysis device concerning an embodiment of the invention. It is a flowchart figure which shows the learning process routine in the signal-analysis apparatus based on embodiment of this invention. It is a flowchart figure which shows the parameter estimation processing routine in the signal-analysis apparatus based on embodiment of this invention. It is a figure which shows an experimental result. It is a figure which shows an experimental result. It is a figure which shows an experimental result. It is a figure which shows an experimental result.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Overview of the Invention>
First, an outline of the present embodiment will be described. In the present embodiment, in order to solve the problem of the above-described SSNMF method, more clues for estimating the speech spectrum can be provided by using not only the spectrum but also the advance information of the feature amount, and distortion of the feature amount is less likely We propose a regularization method of SSNMF.

<Principle of this embodiment>
Next, the principle of the present embodiment will be described.

<Formulation of the problem>
The amplitude spectrogram or power spectrogram of the observation signal (hereinafter, observation spectrogram) is represented by _{Yω, t} . Where ω and t are indexes of frequency and time. Assuming the additivity of the spectrum, the speech spectrum X ^(s) _{ω, t at} each time and the noise spectrum X ⁽ⁿ⁾ _{ω, t} each have K _s basis spectra

And K _n basis spectra

Nonnegative binding of

It can be expressed by

SSNMF method pre-learned from clean speech learning samples

To the observed spectrum Y _{ω, t}

Is a method of estimating the sound component and the noise component contained in the observation spectrogram by fitting. A speech signal can be obtained from the observation signal by a Wiener filter or the like from the speech spectrum and noise spectrum estimated values thus obtained. In this approach, the basis spectrum of pre-learned speech is a clue to the separation of speech and noise, but Y _{ω, t} and X can be used because the noise spectrum can be explained by the basis spectrum of speech and vice versa. _{Even if} the error of _{ω, t} can be reduced, X ^(s) _{ω, t} and X ⁽ⁿ⁾ _{ω, t} do not necessarily correspond to the actual speech spectrum and noise spectrum. For this reason, in order to eliminate the indeterminacy of X ^(s) _{ω, t} and X ⁽ⁿ⁾ _{ω, t} giving the same X _{ω, t} , a stronger constraint that the speech spectrum must satisfy is required. Now, if X 2 ^(s) _{ω, t} corresponds to the speech spectrum, X 2 ^(s) _{ω, t} should be distributed within the range that the speech can actually take in the feature space as well. Therefore, in the present embodiment, paying attention to the cepstrum feature amount, the regularization term for X ^(s) _{ω, t} is considered based on the probability distribution defined in the cepstrum space, and this and Y _{ω, t} and X _ω We propose a parameter optimization algorithm based on the sum of _T , _{and t} error criteria.

The errors of Y _{ω, t} and X _{ω, t} can be measured by squared error, I divergence, Itakura-Satoshi distance, etc., but here I divergence

Use However, all basis spectra are

It is assumed that the following constraints are satisfied. Next, for X ^(s) _{ω, t}

Consider a criterion like However,

Is the Mel-Frequency Cepstrum Coefficients (MFCC) of X _{0, t} , ..., X _{Ω -1, t} , f _{l, ω} is the l-th Mel filter bank coefficient,

Is the coefficient of the discrete cosine transform. Formula (5) is

Is the parameter

Represents the logarithm of the probability generated from the mixed normal distribution of However,

Represents the mean, variance and weight of the mth normal distribution. By learning the parameter θ of this mixed normal distribution from the MFCC series of clean speech learning samples,

Let X ^(s) _{ω, t} be a criterion that gives a high score if it is distributed as much as possible in the MFCC space as the learning samples. The proposed method takes into account the two criteria of equation (3) and equation (5)

The goal is to minimize such criteria. Where λ is a regularization parameter. As described above, this optimization problem is a problem that imposes a soft constraint on the spectrum model by the cepstrum distance criterion, and the inventors have so far used this framework under the same optimization criterion for tone separation and timbre clustering. It proposes the method to do.

  The method of the present embodiment is a method aiming at simultaneously realizing speech signal emphasis and feature quantity emphasis by this framework, and is called "Cepstrum regularization SSNMF".

<Parameter estimation algorithm>

Although it is impossible to analytically obtain U ^(s) , H ⁽ⁿ⁾ and U ⁽ⁿ⁾ to minimize ^N , it is possible to derive an iterative algorithm based on the auxiliary function method to search for the local optimal solution of the optimization problem. it can. In the optimization algorithm of the objective function F (θ) minimization problem by the auxiliary function method, first, an auxiliary variable α is introduced,

Design an auxiliary function F + (θ, α) satisfying If such an auxiliary function can be designed,

When

By alternately repeating the above, it is possible to obtain θ which locally minimizes the objective function F (θ). Below,

Derive the auxiliary function of and the update expression based on it.

In terms of, we use the fact that the negative logarithmic function is a convex function, and use Jensen's inequality

An upper bound function such as However, = ^c represents the equal sign concerning only the term depending on the parameter.

Also H and U are here

And ζ _{k, ω, t} is a variable satisfying ζ _{k, ω, t} 0 0, Σ _k ζ _{k, ω, t} = 1, and the equal sign of equation (8) is

It holds in the case of

next,

Design the upper bound function of As with equation (8), taking advantage of the fact that the negative logarithmic function is a convex function, according to the Jensen inequality

An inequality like this is set up. The equal sign of equation (14) is

It holds in the case of continue

Lead the upper bound function of Because quadratic functions are convex functions, Jensen's inequality

An inequality like this is set up. However,

It is. β _{l, n, m, t} is any positive constant satisfying Σ _l β _{l, n, m, t} = 1,

Is

(16) is the variable that satisfies

It holds in the case of From equations (16) and (14),

It can be said. However,

It is. Note that A _{l, t} is nonnegative and then consider the upper bound function of (log G _{l, t} ) ² . The upper bound of (log G _{l, t} ) ² uses inequality

Can be given by However,

And the equal sign of equation (21) is

It holds in the case of Furthermore, from the fact that f _{l, ω} H _{k, ω} U _{k, t} are nonnegative values, and that the reciprocal function is a convex function in the positive region, according to Jensen's inequality

Is true. However, _{l l, k, ω, t} are variables satisfying を 満 た す_{l, k, ω, t} > 0, _{ω ω, k l l, k, ω, t} = 1,

The equality of equation (26) holds when Then consider the upper bound of the term B _{l, t} log G _{l, t} . Since B _{l, t} is not necessarily a nonnegative value, another inequality is made according to the sign of B _{l, t} . First, since the logarithmic function is a concave function, when B _{l, t} 00,

Get an inequality like φ _{l, t} is a positive variable,

The equality sign of equation (28) holds when On the other hand, when B _{l, t} <0, the negative logarithm function is more convex than the convex function by Jensen's inequality

It can be said. However, _{k k, l, ω, t} are variables satisfying ν _{k, l, ω, t} > 0, Σ _{k, ω k k, l, ω, t} = 1

The equality of equation (30) holds when Summary,



I can write. However, δ _x is an instruction function that is 1 when the condition x is satisfied, and 0 when the condition x is not satisfied. From the above,

Upper bound function of

You can get The point of deriving this inequality is that H and U that minimize the right side can be obtained analytically,

By combining with, it is possible to design an auxiliary function that gives an update expression in closed form. Update formula for each parameter from this auxiliary function

Get However,


It is.

Also, since β _{l, n, m, t} is any positive constant, how is it within the range satisfying β _{l, n, m, t} > 0, _{l l} β _{l, n, m, t} = 1 Even if given, the convergence of the algorithm is guaranteed. Thus, for example, at each step of the iterative calculation

Even if updated, the convergence of the algorithm is guaranteed. This is an update that minimizes the auxiliary function with respect to β _{l, n, m, t} .

<Configuration of Signal Analysis Device According to Embodiment of the Present Invention>
Next, the configuration of the signal analysis device according to the embodiment of the present invention will be described. As shown in FIG. 1, a signal analysis apparatus 100 according to an embodiment of the present invention is a ROM storing a CPU, a RAM, a program for executing a learning process routine and a parameter estimation process routine described later, and various data. And can be configured on a computer. The signal analysis apparatus 100 functionally includes an input unit 10, an operation unit 20, and an output unit 90, as shown in FIG.

  The input unit 10 receives time-series data of a clean audio signal (hereinafter, a clean audio signal) in which other sound sources are not mixed. Further, the input unit 10 receives time-series data of an acoustic signal (hereinafter, an observation signal) in which an audio signal and a noise signal are mixed.

  The operation unit 20 includes a time-frequency expansion unit 24, a feature extraction unit 26, a basis spectrum learning unit 28, a basis spectrum storage unit 30, a feature amount model learning unit 32, a feature amount model storage unit 34, and parameters. An estimation unit 36 and an audio signal generation unit 38 are included.

  The time-frequency expansion unit 24 calculates an amplitude spectrogram or power spectrogram representing time-frequency components of each frequency at each time based on time-series data of the clean speech signal. In the first embodiment, time frequency expansion such as short time Fourier transform or wavelet transform is performed.

In addition, the time frequency expansion unit 24 calculates an observation spectrogram Y which is an amplitude spectrogram or a power spectrogram representing an observation time frequency component Y _{ω, t} of each frequency ω at each time t based on time series data of the observation signal. .

  The feature quantity extraction unit 26 extracts the cepstrum feature quantity of each time based on the time frequency component of each frequency at each time of the clean speech signal calculated by the time frequency expansion unit 24.

Based on the time frequency components of each frequency at each time of the clean speech signal calculated by the time frequency expansion unit 24, the basis spectrum learning unit 28 uses the conventional NMF to calculate each basis k of the clean speech signal. And a base spectrum representing the power spectrum at each frequency ω

Estimate

The basis spectrum storage unit 30 is a basis spectrum representing the power spectrum at each frequency k and each basis k of the clean speech signal estimated by the basis spectrum learning unit 28.

I remember.

  The feature amount model learning unit 32 learns a parameter representing the probability distribution of the cepstrum feature amounts defined in the cepstrum space, based on the cepstrum feature amounts at each time point extracted by the feature amount extraction unit 26. Specifically, when the probability distribution is a mixture normal distribution, the parameter θ of the mixture normal distribution is learned.

  The feature amount model storage unit 34 stores the parameter θ of the mixed normal distribution, which is learned by the feature amount model learning unit 32.

The parameter estimation unit 36 stores the observation spectrogram Y output from the time-frequency expansion unit 24, the basis spectra at each base and each frequency of the audio signal stored in the basis spectrum storage unit 30, and the feature amount model storage unit 34 Each time and each frequency obtained from the observation time frequency component Y of each time and each frequency, the base spectrum of the sound signal, and the activation parameter of the sound signal based on the parameter θ representing the probability distribution of the cepstral feature Of the time frequency component X ^{(s) of} the noise signal and the sum X of the time frequency component X ⁽ⁿ⁾ of each time and each frequency determined from the basis spectrum of the noise signal and the activation parameter

And the likelihood of the cepstral feature of the time-frequency component X ^(s) of each time and each frequency based on the probability distribution of the cepstral feature of the audio signal

The basis spectrum H ^(s) of the speech signal and the activation parameter U ^(s), and the basis spectrum H of the noise signal so as to reduce the criterion shown in the above equation (7) expressed using the regularization term representing ⁽ⁿ⁾ and the activation parameter U ⁽ⁿ⁾ are estimated.

  Specifically, the parameter estimation unit 36 includes an initial value setting unit 40, an auxiliary variable updating unit 42, a parameter updating unit 44, and a convergence determination unit 46.

The initial value setting unit 40 sets, as an initial value of the basis spectrum H ^(s) of the speech signal, basis spectra of each basis and each frequency of the speech signal stored in the basis spectrum storage unit 30. Furthermore, the initial value setting unit 40 sets initial values for the activation parameter U ^(s) of the audio signal, the base spectrum H ^{(n) of} the noise signal, and the activation parameter U ⁽ⁿ⁾ .

The auxiliary variable updating unit 42 uses the parameter θ of the probability distribution of the cepstrum feature stored in the feature amount model storage unit 34 and the basis spectrum H ^{(s) of the} speech signal which is the initial value or updated last time. The above equations (13), (15), (24), and (27) are based on the tivation parameter U ^(s) and the basis spectrum H ^{(n) of} the noise signal and the activation parameter U ^(n). , (29), (31), and (37), each base k, each frequency ω, and ζ _{k, ω, t} for each time _t , each normal distribution m, and α _{m, t} for each time _t , Each mel filter bank coefficient l and ξ _{l, t} , φ _l, t for each time t, each mel filter bank coefficient l, each base k, each frequency ω, and ρ _{l, k, ω, t} for each time _{t , ν k, l, ω,} t, each mel filter bank coefficients l, the normal distribution m, each main Frequency cepstral coefficients n, and beta _{l, m} for each time _{t, n,} and updates the _t.

The parameter updating unit 44 includes the observation spectrogram Y output from the time-frequency expanding unit 24, each base k updated by the auxiliary variable updating unit 42, each frequency ω, and ζ _{k, ω, t} for each time _t , Normal distribution m and α _m, t for each time t, each mel filter bank coefficient l and ξ _{l, t 1,} φ _{l, t} for each time _t , each mel filter bank coefficient l, each base k, each frequency ω, and Ρ _{l, k, ω, t} , _{k k, l, ω,} t for each time t, each mel filter bank coefficient l, each normal distribution m, each mel frequency cepstrum coefficient n, and β _l, for each time t _{Based on m, n, t} , according to the equations (33) to (36), basis spectra H ^(s) _{k, ω} of each basis k and each frequency ω of the speech signal, each basis k and each time t Activation parameters U ^(s) _{k, t,} and for each base k of the noise signal and for each frequency ω The basis spectrum H ⁽ⁿ⁾ _{k, ω,} and each basis k and the activation parameter U ⁽ⁿ⁾ _{k, t at} each time t are updated.

  The convergence determination unit 46 determines whether or not the convergence condition is satisfied, and repeats the updating process in the auxiliary variable updating unit 42 and the updating process in the parameter updating unit 44 until the convergence condition is satisfied.

  As the convergence condition, for example, it can be used that the number of repetitions reaches the upper limit number. Alternatively, as the convergence condition, it can be used that the difference between the value of the criterion of the equation (7) and the value of the previous criterion is equal to or less than a predetermined threshold.

The speech signal generation unit 38 generates the basis spectrum H ^(s) _{k, ω} of each base k and each frequency ω of the speech signal acquired in the parameter estimation unit 36, the basis spectrum H of each base k of each noise signal, and each frequency ω ⁽ⁿ⁾ Based on _{k, ω} , according to the Wiener filter, an audio signal is generated and output from the output unit 90.

<Operation of Signal Analysis Device According to Embodiment of the Present Invention>
Next, the operation of the signal analysis device 100 according to the embodiment of the present invention will be described. First, when time-series data of a clean speech signal is received at the input unit 10, the signal analysis device 100 executes a learning processing routine shown in FIG.

  First, in step S100, based on the time-series data of the clean speech signal received by the input unit 10, the time frequency component of each frequency at each time of the clean speech signal is calculated.

  Next, in step S102, the cepstrum feature value of each time is extracted based on the time frequency component of each frequency at each time of the clean speech signal acquired in step S100.

Next, in step S104, based on the time frequency components of each frequency at each time of the clean speech signal acquired in step S100, the NMF according to the prior art calculates the power at each base k and each frequency ω of the clean speech signal. Basis spectrum representing the spectrum

Are estimated and stored in the basis spectrum storage unit 30.

  In step S106, based on the cepstrum feature quantity at each time point extracted in step S102, the parameter θ of the mixture normal distribution representing the probability distribution of the cepstrum feature quantity is learned, and stored in the feature quantity model storage unit 34. End the processing routine.

  Next, when the input unit 10 receives time-series data of an observation signal in which an audio signal and a noise signal are mixed, the signal analysis device 100 executes a parameter estimation processing routine shown in FIG. 3.

  First, in step S120, the observation spectrogram Y is calculated based on the time-series data of the observation signal received by the input unit 10.

In step S122, as the initial value of the basis spectrum H ^(s) of the speech signal, the basis spectrum of each basis and each frequency of the speech signal stored in the basis spectrum storage unit 30 is set. Further, initial values are set to the activation parameter U ^{(s) of} the audio signal, the base spectrum H ^{(n) of} the noise signal, and the activation parameter U ⁽ⁿ⁾ .

In step S124, the parameter θ of the probability distribution of the cepstrum feature stored in the feature quantity model storage unit 34 and the initial value set in step S122, or of the audio signal updated last in step S126 described later Based on the basis spectrum H ^(s) and the activation parameter U ^(s), and the basis spectrum H ^{(n) of} the noise signal and the activation parameter U ⁽ⁿ⁾ , 24), (27), (29), (31) and (37), each base k, each frequency ω, and ζ _{k, ω, t} for each time _t , each normal distribution m and Α _m, t for each time t, each mel filter bank coefficient l and ξ _{l, t} , φ _{l, t} for each time _t , each mel filter bank coefficient l, each base k, each frequency ω, and each time t _{l l, k, ω, t} , _{k k, 1. l, ω, t} , each mel filter bank coefficient l, each normal distribution m, each mel frequency cepstrum coefficient n, and β _{l, m, n, t} for each time t.

Next, in step S126, the observed spectrogram Y obtained in step S120, each base k updated in step S124, each frequency ω, and ζ _{k, ω, t} for each time _t , each normal distribution m And α _m, t for each time t, each mel filter bank coefficient l and ξ _{l, t} , φ _{l, t} for each time _t , each mel filter bank coefficient l, each base k, each frequency ω, and each time t Ρ _{l, k, ω, t} , _{k k, l, ω, t} , each mel filter bank coefficient l, each normal distribution m, each mel frequency cepstrum coefficient n, and β _{l, m, n,} for each time t Based on _t , according to the equations (33) to (36), the basis spectrum H ^(s) _{k, ω} of each base k and each frequency ω of the speech signal, and the activation parameters of each base k and each time t U ^(s) _{k, t} , the basis sp of each base k of the noise signal and each frequency ω Vector H ⁽ⁿ⁾ _k, ω and activation parameters U ⁽ⁿ⁾ _k of each basis k and each time _t, to update the _t.

  Next, in step S128, it is determined whether the convergence condition is satisfied. If the convergence condition is satisfied, the process proceeds to step S130. If the convergence condition is not satisfied, the process proceeds to step S124, and the processes of steps S124 to S128 are repeated.

In step S130, the basis spectrum H ^(s) _{k, ω} of each base k and each frequency ω of the speech signal finally updated in step S126, and the basis spectrum H of each base k of each noise signal and each frequency ω ⁽ⁿ⁾ Generate an audio signal according to the Wiener filter based on _{k, ω} , and output it from the output unit 90, and complete the parameter estimation processing routine.

<Example of experiment>
Evaluation experiment to verify the noise suppression effect by the method of the above-mentioned embodiment using speech data of ATR speech database 503 sentences and noise data of RWCP (white noise, babble noise, museum noise, background music noise) went. The comparison target was the conventional SSNMF method, and the improvement values of signal-to-distortion ratio (SDR) and cepstrum distortion before and after processing were evaluated. Test data was created by superimposing each noise on clean speech at various SNRs. All acoustic signals in the test data were monaural signals with a sampling frequency of 16 kHz, and a short Fourier transform was performed with a frame length of 32 ms and a frame shift of 16 ms to calculate the observation spectrogram Y _{ω, t} . In the learning, H ^(s) _{k, ω} and the MFCC GMM parameter θ were learned using speech of a total of 450 sentences of 10 speakers (including 4 females and 6 males). The dimension of MFCC is 13 and the number of mixed GMMs is 30. In the test, H ^(s) _{k, ω} and θ obtained by learning are fixed, and λ ⁽ = 1 ⁾ _, U ^(s) _{k, t} , H ⁽ⁿ⁾ _{k, ω} , U ⁽ⁿ⁾ _{k, t} The estimation of After estimation, the estimated value of the speech signal was calculated by the Wiener filter using X ^(s) _{ω, t} and X ⁽ⁿ⁾ _{ω, t} . The initial values of the proposed algorithm were obtained by the conventional SSNMF.

  The improved values of cepstrum distortion and SDR obtained by the proposed method and the conventional method under the above conditions are shown in FIGS.

  FIG. 4 shows the improvement of cepstrum distortion obtained by the proposed method and the conventional method. The upper part of FIG. 4 shows the case where the type of noise is white noise, and the lower part of FIG. 4 shows the case where the type of noise is bubble noise. FIG. 5 shows the improvement of cepstrum distortion obtained by the proposed method and the conventional method. The upper part of FIG. 5 shows the case where the type of noise is real environment noise, and the lower part of FIG. 5 shows the case where the type of noise is background music noise. FIG. 6 shows the improvement value of SDR obtained by the proposed method and the conventional method. The upper part of FIG. 6 shows the case where the type of noise is white noise, and the lower part of FIG. 6 shows the case where the type of noise is bubble noise. FIG. 7 shows the improvement value of SDR obtained by the proposed method and the conventional method. The upper part of FIG. 7 shows the case where the type of noise is real environment noise, and the lower part of FIG. 7 shows the case where the type of noise is background music noise. It was confirmed that the proposed method achieved higher improvement values in most cases in any of the evaluation scales.

  As described above, according to the signal analysis device according to the embodiment of the present invention, the observation time frequency components of each time and each frequency, the base spectrum of the speech signal estimated in advance, and the activation parameter of the speech signal Distance with the sum of time-frequency components of each time and frequency determined from time-frequency components of each time and frequency determined from the time base and activation parameters of noise signal, and probability of cepstral feature of speech signal In order to reduce the criterion expressed using a regularization term that represents the likelihood of the cepstral feature of the time-frequency component of each time and each frequency determined from the basis spectrum and activation parameters of the speech signal based on the distribution, Voice signal activation parameters and noise signal basis Estimating the torque and the activation parameters, by generating an audio signal to suppress noise, as well as emphasizing the speech signal, it can be emphasized cepstrum characteristic quantity.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made without departing from the scope of the present invention.

  For example, it is configured to perform processing of learning parameters representing probability distribution of basis spectrum of speech signal and cepstrum feature of speech signal and parameter estimation for estimating basis spectrum and activation parameter from observation signal by different devices. You may

  In addition, since the order of parameters to be updated is arbitrary, it is not limited to the order of the above embodiment.

  Also, although the case of updating the base spectrum of the audio signal as well as the activation parameter of the audio signal, the base spectrum of the noise signal, and the activation parameter has been described as an example, the present invention is not limited thereto. The base spectrum of the speech signal may be fixed to the base spectrum of the pre-estimated speech signal without updating the base spectrum of.

  Furthermore, although the present invention has been described as an embodiment in which the program is installed in advance, the program can be provided by being stored in a computer readable recording medium, and provided via a network. It is also possible.

10 input unit 20 operation unit 24 time frequency expansion unit 26 feature amount extraction unit 28 basis spectrum learning unit 30 basis spectrum storage unit 32 feature amount model learning unit 34 feature amount model storage unit 36 parameter estimation unit 38 speech signal generation unit 40 initial value Setting unit 42 Auxiliary variable updating unit 44 Parameter updating unit 46 Convergence determination unit 90 Output unit 100 Signal analysis device

Claims (7)

A time-frequency expansion unit that receives, as input, time-series data of an observation signal in which an audio signal and a noise signal are mixed, and outputs an observation spectrogram representing an observation time-frequency component of each time and each frequency;
Cepstrum features of a pre-learned speech signal defined by the observation spectrogram output by the time-frequency expansion unit, a basis spectrum representing a power spectrum at each basis and frequency of each basis and frequency of the previously learned speech signal, and a cepstrum space Based on the parameters representing the probability distribution of quantities,
Time frequency components of each time and each frequency obtained from activation time parameters indicating observation time frequency components of each time and each frequency, base spectrum of the sound signal, and power at each time of the sound signal, and the noise signal The base spectrum of the audio signal based on the base spectrum and the distance between the time and the sum of time frequency components of each frequency obtained from the activation parameter, and the probability distribution of the cepstral feature of the audio signal; The activation parameters of the audio signal so as to reduce the criterion expressed using a regularization term representing the likelihood of the cepstral feature of the time-frequency component of each time and each frequency obtained from the activation parameter; The basis of the noise signal A parameter estimation unit for estimating the torque and the activation parameter,
Signal analyzer including: The signal analysis apparatus according to claim 1, wherein the criterion is expressed by the following equation.


Where U ^(s) represents the activation parameter of the speech signal, H ⁽ⁿ⁾ represents the base spectrum of the noise signal, and U ⁽ⁿ⁾ represents the activation parameter of the noise signal. Y represents the observation spectrogram, and X represents the time-frequency component X ^{(s) of} each frequency determined from the basis spectrum of the speech signal and the activation parameter U ^(s), and the basis of the noise signal Represents the sum of time frequency components of each frequency obtained from the spectrum H ⁽ⁿ⁾ and the activation parameter U ⁽ⁿ⁾ ,


Represents the distance between the observed spectrogram Y and the sum X,


Represents the likelihood of the cepstral feature of the time-frequency component X ^(s) of each frequency determined from the basis spectrum of the voice signal and the activation parameter U ^(s) , and λ is a regularization parameter. The parameter estimation unit
So as to reduce the auxiliary function is the upper bound function of the criterion, the parameter updating section for updating said activation parameter of the audio signal, and a pre-Symbol basal spectrum and the activation parameters of the noise signal,
A convergence determination unit that repeats updating by the parameter updating unit until a predetermined convergence condition is satisfied;
The signal analysis device according to claim 1 or 2, comprising A signal analysis method in a signal analysis apparatus including a time frequency expansion unit and a parameter estimation unit,
The time-frequency expansion unit receives time-series data of an observation signal in which an audio signal and a noise signal are mixed, and outputs an observation spectrogram representing an observation time-frequency component of each time and each frequency,
The parameter estimation unit is pre-learned defined by the observation spectrogram output by the time-frequency expansion unit, a basis spectrum representing a power spectrum at each basis and each frequency of each basis and a frequency of a speech signal learned in advance, and a cepstrum space Based on the parameters representing the probability distribution of the cepstral feature of the speech signal
Time frequency components of each time and each frequency obtained from activation time parameters indicating observation time frequency components of each time and each frequency, base spectrum of the sound signal, and power at each time of the sound signal, and the noise signal The base spectrum of the audio signal based on the base spectrum and the distance between the time and the sum of time frequency components of each frequency obtained from the activation parameter, and the probability distribution of the cepstral feature of the audio signal; The activation parameters of the audio signal so as to reduce the criterion expressed using a regularization term representing the likelihood of the cepstral feature of the time-frequency component of each time and each frequency obtained from the activation parameter; The basis of the noise signal Torr and signal analysis method for estimating and said activation parameters. The signal analysis method according to claim 4, wherein the criterion is expressed by the following equation.


Where U ^(s) represents the activation parameter of the speech signal, H ⁽ⁿ⁾ represents the base spectrum of the noise signal, and U ⁽ⁿ⁾ represents the activation parameter of the noise signal. Y represents the observation spectrogram, and X represents the time-frequency component X ^{(s) of} each frequency determined from the base spectrum of the speech signal and the activation parameter U ^(s), and the noise signal Represents the sum of time frequency components of each frequency obtained from the basis spectrum H ⁽ⁿ⁾ and the activation parameter U ⁽ⁿ⁾ ,


Represents the distance between the observed spectrogram Y and the sum X,


Represents the likelihood of the cepstral feature of the time-frequency component X ^(s) of each frequency determined from the basis spectrum of the voice signal and the activation parameter U ^(s) , and λ is a regularization parameter. In the estimation by the parameter estimation unit,
Parameter updating unit, to reduce the auxiliary function is the upper bound function of the criterion, then updating said activation parameter of the audio signal, and said base spectrum and the activation parameters of the noise signal,
The signal analysis method according to claim 4 or 5, further comprising repeating the updating by the parameter updating unit until the convergence determination unit satisfies a predetermined convergence condition.   The program for functioning a computer as each part of the signal-analysis apparatus in any one of Claims 1-3.