JP2017151222A - Signal analysis device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To suppress a noise, to emphasize a voice signal, and to emphasize cepstrum feature quantity.SOLUTION: A parameter estimation part 36 estimates an activation parameter of a voice signal and a base spectrum and activation parameter of a noise signal to reduce a standard represented by using a distance between an observation time frequency component and a sum of a time frequency component calculated from a preliminarily estimated base spectrum of the voice signal and the activation parameter of the voice signal and a time frequency component calculated from the base spectrum and activation parameter of the noise signal and a regularization item representing the likelihood of cepstrum feature quantity of a time frequency component calculated from the base spectrum and activation parameter of the voice signal based on probability distribution of cepstrum feature quantity of the voice signal. A voice signal generation part 38 generates the voice signal from the base spectrums of the voice signal and the noise signal.SELECTED DRAWING: Figure 1

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. Noise mixed in an audio signal not only deteriorates the quality of audio communication, but also causes a reduction in performance of various audio processing such as audio recognition and audio conversion. In order to solve this problem, various speech enhancement methods have been proposed so far.

  Speech enhancement methods are broadly divided into unsupervised approaches, supervised approaches, and semi-supervised approaches.

  In the supervised approach, the target speech and target noise samples are obtained in advance, in the semi-supervised approach, only the target speech sample is obtained in advance, and in the unsupervised approach, both are assumed. Emphasis technique. Moreover, it is divided roughly into the case where the object to emphasize is a signal (or spectrum), and the case of a feature-value. Typical examples of supervised feature enhancement approaches include Vector Taylor Series (VTS), Stereo Piecewise Linear Compensation for Environment (SPLICE), and Denoising Autoencoder (DAE).

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

  SPLICE uses a mixed normal distribution (Gaussian Mixture Model: GMM) to model the joint probability density function of features with noise and clean speech, and clean speech from noise features with GMM parameters learned using training samples. This is a technique for constructing a conversion function to a feature quantity.

  The DAE method is a method of constructing a conversion function between input and output by a deep neural network having a speech feature with noise as an input and a clean speech feature as an output. These supervised speech enhancement approaches are extremely powerful in known noise environments because they are based on discriminative models and discriminative criteria, but are not necessarily effective in unknown noise environments. However, many methods have been proposed to compensate for the mismatch when the voice or noise of the learning data is different from that at the time of the test.

  On the other hand, a method based on semi-supervised non-negative matrix factorization (SSNMF), which is a representative example of a semi-supervised signal enhancement approach, is performed in an unknown noise environment. In recent years, it has attracted attention 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 observed spectrum at each time with a non-negative combination of the base spectrum of speech and the base spectrum of noise.

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

  However, if the noise spectrum can be explained by the speech base spectrum, or vice versa, the estimated spectrum may not correspond to the actual speech spectrum. For this reason, in order to eliminate the indefiniteness of the decomposition of the speech spectrum and the noise spectrum, stronger constraints that the speech spectrum should satisfy are required. In addition, even if the signal can be enhanced in the SSNMF method, there is no guarantee that the feature amount can be enhanced. Therefore, the enhancement processing does not necessarily directly improve the performance of speech processing based on speech feature amounts such as speech recognition and speech conversion.

  The present invention has been made in view of the above circumstances, and provides a signal analysis apparatus, method, and program capable of suppressing noise, enhancing a speech signal, and enhancing a cepstrum feature amount. Objective.

  In order to achieve the above object, the signal analyzing apparatus according to the present invention receives time series data of an observation signal in which a speech signal and a noise signal are mixed as input, and represents an observation time frequency component at each time and each frequency. A time-frequency expansion unit that outputs a spectrogram, the observation spectrogram output by the time-frequency expansion unit, a base spectrum that represents a power spectrum at each base and each frequency of a previously learned speech signal, and a cepstrum space , Based on the parameters representing the probability distribution of the cepstrum feature amount of the speech signal learned in advance, the observed time frequency component of each time and each frequency, the base spectrum of the speech signal, and the power of the speech signal at each time Each time and frequency period obtained from the activation parameter The speech based on the number component, the distance from the base frequency of the noise signal and the sum of the time frequency components of each time and each frequency obtained from the activation parameter, and the probability distribution of the cepstrum feature of the speech signal The speech signal so as to reduce the criterion expressed using the regularization term representing the likelihood of the cepstrum feature quantity of the time frequency component of each time and each frequency obtained from the base spectrum of the signal and the activation parameter. And the parameter estimation unit for estimating the base spectrum of the noise signal and the activation parameter.

  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. Using the time-series data of the signal as an input, output an observation spectrogram representing the observation time frequency component of each time and frequency, and the parameter estimation unit is trained in advance, the observation spectrogram output by the time frequency expansion unit Based on the base spectrum representing the power spectrum at each base and each frequency of the speech signal and the parameter representing the probability distribution of the cepstrum feature amount of the speech signal learned in advance defined in the cepstrum space, at each time and each frequency. Observation time frequency component, base spectrum of the audio signal, and each time of the audio signal The time frequency component of each time and each frequency obtained from the activation parameter representing the power in the signal, and the distance from the sum of the time frequency component of each time and each frequency obtained from the base spectrum and the activation parameter of the noise signal And regularization representing the likelihood of the time frequency component of each time and frequency obtained from the base spectrum of the speech signal and the activation parameter based on the probability distribution of the cepstrum feature amount of the speech signal The activation parameter of the speech signal, the base spectrum of the noise signal, and the activation parameter are estimated so as to reduce a criterion expressed using a term.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said signal analysis apparatus.

  As described above, according to the signal analysis apparatus, method, and program of the present invention, the observation time frequency component of each time and frequency, the base spectrum of the audio signal, and the power of the audio signal at each time are calculated. A time frequency component of each time and each frequency obtained from the activation parameter to represent, a distance from a sum of the time frequency component of each time and each frequency obtained from the base spectrum of the noise signal and the activation parameter, and Based on the probability distribution of the cepstrum feature amount of the speech signal, using a regularization term representing the likelihood of the cepstrum feature amount of the time frequency component of each time and each frequency obtained from the base spectrum and the activation parameter of the speech signal 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 which shows the functional structure of the signal analyzer which concerns on embodiment of this invention. It is a flowchart figure which shows the learning process routine in the signal analyzer which concerns on embodiment of this invention. It is a flowchart figure which shows the parameter estimation processing routine in the signal analyzer which concerns 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.

<Outline of the invention>
First, an outline of the present embodiment will be described. In the present embodiment, in order to solve the above-described problem of the SSNMF method, not only the spectrum but also the prior information of the feature amount is used to give more clues for estimating the speech spectrum and to prevent distortion of the feature amount. We propose a regularization method for SSNMF.

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

<Formulation of problem>
Let Y _{ω, t} be the amplitude spectrogram or power spectrogram (hereinafter referred to as observation spectrogram) of the observed signal. Where ω and t are frequency and time indexes. Assuming spectrum additivity, K _s basis spectra of speech spectrum X ^(s) _{ω, t} and noise spectrum X ⁽ⁿ⁾ _{ω, t} at each time

And K _n basis spectra

Non-negative coupling of

It can be expressed as

The SSNMF method was pre-trained from a clean speech training sample.

To the observed spectrum Y _{ω, t}

This is a method for estimating speech components and noise components included in an observation spectrogram by fitting. A speech signal can be obtained from the observation signal by a Wiener filter or the like from the estimated values of the speech spectrum and the noise spectrum obtained in this way. In this approach, the pre-learned speech base spectrum is a clue to separation of speech and noise, but the noise spectrum can be explained by the speech base spectrum and vice versa, so Y _{ω, t} and X _{Even if} the error of _{ω, t} can be reduced, X ^(s) _{ω, t} and X ⁽ⁿ⁾ _{ω, t} do not always correspond to the actual speech spectrum and noise spectrum. For this reason, in order to eliminate the indefiniteness of X ^(s) _{ω, t} and X ⁽ⁿ⁾ _{ω, t} giving the same X _{ω, t} , a stronger constraint that the speech spectrum should satisfy is necessary. Now, if X ^(s) _{ω, t} corresponds to the speech spectrum, X ^(s) _{ω, t} should be distributed within the range that the speech can actually take in the feature amount space. Therefore, in this embodiment, paying attention to the cepstrum feature amount, a regularization term for X ^(s) _{ω, t} is considered based on the probability distribution defined in the cepstrum space, and this is _{expressed as} Y _{ω, t} and X _ω Therefore, a parameter optimization algorithm based on the sum of _t error criterion is proposed.

The error between Y _{ω, t} and X _{ω, t} can be measured by square error, I divergence, Itakura Saito distance, etc., but here I divergence

Is used. However, all base spectra are

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

Consider the following criteria. 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 a coefficient of discrete cosine transform. Equation (5) is

Is a parameter

Represents the logarithm of the probability generated from the mixed normal distribution. 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 sequence of the clean speech learning sample,

Can be a criterion that gives a high score when X ^(s) _{ω, t} is distributed in the MFCC space as much as possible in the learning sample. The proposed method considers 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. Until now, the inventors have used this framework to separate musical tone separation and timbre clustering under the same optimization criterion. Proposes a technique to do.

  The method of the present embodiment is a method aiming to simultaneously realize speech signal enhancement and feature amount enhancement by this framework, and is called “cepstrum regularization SSNMF”.

<Parameter estimation algorithm>

U ^(s) , H ⁽ⁿ⁾ , U ⁽ⁿ⁾ cannot be obtained analytically, but an iterative algorithm that searches for the local optimal solution of the optimization problem can be derived based on the auxiliary function method it can. In the optimization algorithm for the minimization problem of the objective function F (θ) by the auxiliary function method, first, the auxiliary variable α is introduced,

An auxiliary function F + (θ, α) that satisfies the above is designed. If such an auxiliary function can be designed,

When

By alternately repeating the above, θ that locally minimizes the objective function F (θ) can be obtained. Below,

Auxiliary functions and update formulas based on them are derived.

For, the negative logarithmic function is a convex function, and Jensen's inequality

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

H and U here

It is said. ζ _{k, ω, t} is a variable that satisfies ζ _{k, ω, t} ≧ 0 and Σkζ _{k, ω, t} = 1.

This holds true.

next,

Design the upper bound function of. Similar to Eq. (8), the negative logarithmic function is a convex function.

An inequality such as The equal sign in equation (14) is

This holds true. continue

Deriving the upper bound function of Since the quadratic function is a convex function, from Jensen's inequality

An inequality such as However,

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

Is

And the equal sign in equation (16) is

This holds true. From Equation (16) and Equation (14),

I can say. 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 is

Can be given in However,

And the equal sign in equation (21) is

This holds true. Furthermore, since f _{l, ω} H _{k, ω} U _{k, t} are non-negative values and the reciprocal function is a convex function in the positive region, Jensen's inequality is

Holds. However, ρ _{l, k, ω, t} is a variable satisfying ρ _{l, k, ω, t} > 0 and Σ _{ω, k} ρ _{l, k, ω, t} = 1,

Then the equal sign of equation (26) holds. Next, consider the upper bound of the term B _{l, t} log G _{l, t} . Since B _{l, t} is not necessarily a non-negative value, another kind of inequality is established according to the sign of B _{l, t} . First, since the logarithmic function is a concave function, when B _{l, t} ≧ 0,

To obtain an inequality such as φ _{l, t} is a positive variable,

Then the equal sign of equation (28) holds. On the other hand, when B _{l, t} <0, the negative logarithmic function is more

I can say. Where ν _{k, l, ω, t} is a variable satisfying ν _{k, l, ω, t} > 0 and Σ _{k, ω} ν _{k, l, ω, t} = 1,

Then the equal sign of equation (30) holds. Summary,



Can be written. However, δ _x is an indicator function that becomes 1 when the condition x is satisfied and becomes 0 when the condition x is not satisfied. From the above,

Upper bound function

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

It is possible to design an auxiliary function that gives an update expression in a closed form by combining with. From this auxiliary function, the update formula for each parameter

Get. However,


It is.

Also, since β _{l, n, m, t} is an arbitrary positive constant, how is it as long as β _{l, n, m, t} > 0 and Σ _l β _{l, n, m, t} = 1 are satisfied? Even if given, the convergence of the algorithm is guaranteed. So, 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 apparatus according to the embodiment of the present invention will be described. As shown in FIG. 1, a signal analyzing apparatus 100 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing a learning processing routine and a parameter estimation processing routine described later. And a computer including the above. Functionally, the signal analyzing apparatus 100 includes an input unit 10, an arithmetic 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, clean audio signal) that is not mixed with other sound sources. The input unit 10 also receives time-series data of an acoustic signal (hereinafter referred to as an observation signal) in which an audio signal and a noise signal are mixed.

  The calculation unit 20 includes a time frequency expansion unit 24, a feature amount extraction unit 26, a base spectrum learning unit 28, a base 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 a power spectrogram representing a time frequency component of each frequency at each time based on the time series data of the clean speech signal. In the first embodiment, time-frequency expansion such as short-time Fourier transform and wavelet transform is performed.

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

  The feature amount extraction unit 26 extracts the cepstrum feature amount at 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.

The base spectrum learning unit 28 uses the conventional NMF to calculate each basis k of the clean speech signal 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. And the base spectrum representing the power spectrum at each frequency ω

Is estimated.

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

Is remembered.

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

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

The parameter estimation unit 36 is stored in the observation spectrogram Y output from the time-frequency expansion unit 24, each base of the speech signal stored in the base spectrum storage unit 30, and the base spectrum at each frequency, 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 audio signal, and the activation parameter of the audio signal based on the parameter θ representing the probability distribution of the cepstrum feature quantity And the time frequency component X ^{(s) of} the noise signal and the distance X from the sum X of the time frequency component X ⁽ⁿ⁾ of each time and each frequency obtained from the base spectrum of the noise signal and the activation parameter

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

The base spectrum H ^(s) and the activation parameter U ^(s) of the speech signal and the base 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 update unit 42, a parameter update unit 44, and a convergence determination unit 46.

The initial value setting unit 40 sets a base spectrum in each base and each frequency of the speech signal stored in the base spectrum storage unit 30 as an initial value of the base spectrum H ^(s) of the speech signal. 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 includes the cepstrum feature quantity probability distribution parameter θ stored in the feature quantity model storage unit 34, the initial value, or the previously updated base spectrum H ^(s) and the active spectrum of the speech signal. Based on the activation parameter U ^(s) , the base spectrum H ^{(n) of} the noise signal, and the activation parameter U ⁽ⁿ⁾ , the above expressions (13), (15), (24), and (27) , (29), (31), and (37), ζ _{k, ω, t} for each basis k, each frequency ω, and each time _t , each normal distribution m, and α _{m, t} for each time _t , Ξ _{l, t} , φ _l, t for each mel filter bank coefficient l and 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 update unit 44 includes the observation spectrogram Y output by the time frequency expansion unit 24, each base k updated by the auxiliary variable update unit 42, each frequency ω, and ζ _{k, ω, t} for each time _t , Α _m, t for each normal distribution m and 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 basis k, each frequency ω, and Ρ _{l, k, ω, t} , ν _{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} , the base spectrum H ^(s) _{k, ω} of each base k and each frequency ω of the speech signal, each base k, and each time t according to the above formulas (33) to (36). Activation parameter U ^(s) _{k, t} of each noise signal basis k and each frequency ω Update the base spectrum H ⁽ⁿ⁾ _{k, ω} and the activation parameters U ⁽ⁿ⁾ _{k, t} for each base k and each time t.

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

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

The sound signal generation unit 38 has a base spectrum H ^(s) _{k, ω} of each base k and each frequency ω of the sound signal acquired by the parameter estimation unit 36, and each base k and base spectrum H of each frequency ω of the aligned noise signal. ⁽ⁿ⁾ Based on _{k and ω} , an audio signal is generated according to the Wiener filter 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 analyzing apparatus 100 according to the embodiment of the present invention will be described. First, when the time series data of the clean speech signal is received at the input unit 10, the signal analyzing apparatus 100 executes a learning processing routine shown in FIG.

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

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

Next, in step S104, based on the time frequency component of each frequency at each time of the clean speech signal acquired in step S100, the power at each base k and each frequency ω of the clean speech signal is obtained by NMF, which is a conventional technique. The base spectrum representing the spectrum

Is stored in the base spectrum storage unit 30.

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

  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 apparatus 100 executes a parameter estimation processing routine shown in FIG.

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

In step S122, the base spectrum at each base and each frequency of the speech signal stored in the base spectrum storage unit 30 is set as the initial value of the base spectrum H ^(s) of the speech signal. Also, initial values are set 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⁾ .

In step S124, the cepstrum feature quantity probability distribution parameter θ stored in the feature quantity model storage unit 34 and the initial value set in step S122 or updated last time in step S126 described later are used. Based on the base spectrum H ^(s) and the activation parameter U ^(s) and the base spectrum H ⁽ⁿ⁾ and the activation parameter U ^{(n) of} the noise signal, the above formulas (13), (15), ( 24), (27), (29), (31), and (37), ζ _{k, ω, t} for each base k, each frequency ω, and 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} , ν _{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, t} for each time _t are updated.

Next, in step S126, the observation spectrogram Y obtained in step S120, each basis k updated in step S124, each frequency ω, and ζ _{k, ω, t} for each time t, and each normal distribution m. Α _m, t for each time t, each mel filter bank coefficient l and ξ _{l, t} , φ _{l, t} , each mel filter bank coefficient l, each base k, each frequency ω, and each time t for each time t Ρ _{l, k, ω, t} , ν _{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, On} the basis of _t , according to the above formulas (33) to (36), the basis spectrum H ^(s) _{k, ω} for each base k and each frequency ω of the speech signal, the activation parameters for each base k and each time t U ^(s) _{k, t} and the basis spectrum of each basis k and frequency ω of the noise signal Updates the spectrum H ⁽ⁿ⁾ _{k, ω} and the activation parameters U ⁽ⁿ⁾ _{k, t} for each basis k and each time t.

  Next, in step S128, it is determined whether or not a 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 in steps S124 to S128 are repeated.

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

<Experimental example>
ATR speech database An evaluation experiment to verify the noise suppression effect by the method of the above embodiment using speech data of 503 sentences and RWCP noise data (four types of white noise, babble noise, museum noise, and background music noise) went. The comparison target was the conventional SSNMF method, and the improvement in signal-to-distortion ratio (SDR) and cepstrum distortion before and after processing were evaluated. The test data was created by superimposing each noise on clean speech with various SNRs. All acoustic signals in the test data were monaural signals with a sampling frequency of 16 kHz, and Fourier transform was performed for a short time with a frame length of 32 ms and a frame shift of 16 ms, and the observation spectrogram Y _{ω, t} was calculated. In learning, H ^(s) _{k, ω} and MFCC GMM parameter θ were trained using a total of 450 sentences of 10 speakers (including 4 women and 6 men). The MFCC dimension was 13 and the number of GMM mixtures was 30. In the test, H ^(s) _{k, ω} and θ obtained by learning are fixed, and U ^(s) _{k, t} , H ⁽ⁿ⁾ _{k, ω} , U ⁽ⁿ⁾ _{k, t} with λ = 1 Was estimated. 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 value of the proposed algorithm was obtained by 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 value of the cepstrum distortion obtained by the proposed method and the conventional method. 4 shows the case where the noise type is white noise, and the lower part of FIG. 4 shows the case where the noise type is bubble noise. FIG. 5 shows the improvement value of the cepstrum distortion obtained by the proposed method and the conventional method. FIG. 5 shows the case where the noise type is real environment noise, and FIG. 5 shows the case where the noise type is background music noise. FIG. 6 shows the improved SDR values obtained by the proposed method and the conventional method. 6 shows the case where the noise type is white noise, and the lower part of FIG. 6 shows the case where the noise type is bubble noise. FIG. 7 shows the improved SDR values obtained by the proposed method and the conventional method. 7 shows the case where the noise type is real environment noise, and the lower part of FIG. 7 shows the case where the noise type is background music noise. In any evaluation scale, it was confirmed that the proposed method was able to obtain higher improvement values in most cases.

  As described above, according to the signal analysis device according to the embodiment of the present invention, the observation time frequency component of each time and each frequency, the base spectrum of the speech signal estimated in advance, and the activation parameter of the speech signal The time frequency component of each time and frequency obtained from the above, the distance from the sum of the time frequency components of each time and frequency obtained from the base spectrum and activation parameters of the noise signal, and the probability of the cepstrum feature quantity of the audio signal Based on the distribution, the criterion expressed using the regularization term representing the likelihood of the cepstrum feature amount of the time-frequency component of each time and each frequency obtained from the base spectrum and the activation parameter of the audio signal is reduced. Audio signal activation parameters and noise signal basis spectrum 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.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, the processing for learning the parameters representing the probability distribution of the speech signal base spectrum and the speech signal cepstrum feature amount and the parameter estimation for estimating the base spectrum and the activation parameter from the observation signal are performed by different devices. May be.

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

  In addition, the case where the base spectrum of the audio signal is updated 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, but the present invention is not limited to this. The base spectrum of the speech signal may be fixed to the preestimated speech spectrum without updating the base spectrum.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do.

DESCRIPTION OF SYMBOLS 10 Input part 20 Operation part 24 Time frequency expansion part 26 Feature-value extraction part 28 Base spectrum learning part 30 Base spectrum memory | storage part 32 Feature-value model learning part 34 Feature-value model memory | storage part 36 Parameter estimation part 38 Speech signal generation part 40 Initial value Setting unit 42 Auxiliary variable update unit 44 Parameter update unit 46 Convergence determination unit 90 Output unit 100 Signal analysis device

Claims (7)

A time-frequency expansion unit that outputs an observation spectrogram representing an observation time-frequency component of each time and each frequency, using as input the time-series data of the observation signal in which the audio signal and the noise signal are mixed,
The observation spectrogram output by the time-frequency expansion unit, the base spectrum representing the power spectrum at each base and each frequency of the pre-learned speech signal, and the cepstrum feature of the pre-learned speech signal defined by the cepstrum space Based on the parameters representing the probability distribution of quantities,
Observation time frequency components of each time and frequency, the base spectrum of the audio signal, and the time frequency component of each time and frequency obtained from an activation parameter representing the power of the audio signal at each time, and the noise signal The base spectrum of the speech signal based on the distance from the sum of the time frequency components of each time and each frequency obtained from the base spectrum and the activation parameter of the speech signal, and the probability distribution of the cepstrum feature amount of the speech signal The activation parameter of the audio signal so as to reduce the criterion expressed using the regularization term representing the likelihood of the cepstrum feature quantity of the time-frequency component of each time and each frequency obtained from the activation parameter; The basis spectrum of the noise signal; A parameter estimation unit for estimating the torque and the activation parameter,
Including a signal analysis device. The signal analysis apparatus according to claim 1, wherein the criterion is represented by the following expression.

Where U ^(s) represents the activation parameter of the audio signal, H ⁽ⁿ⁾ represents the base spectrum of the noise signal, and U ⁽ⁿ⁾ represents the activation parameter of the noise signal. Y is the observation spectrogram, X is the time frequency component X ^{(s) of} each frequency obtained from the base spectrum of the speech signal and the activation parameter U ^(s), and the base of the noise signal It 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 cepstrum feature quantity of the time-frequency component X ^(s) of each frequency obtained from the base spectrum of the speech signal and the activation parameter U ^(s) , and λ is a regularization parameter. The parameter estimation unit includes:
A parameter updating unit that updates the activation parameter of the speech signal and the base spectrum and the activation parameter of the speech signal of the noise signal so as to reduce an auxiliary function that is an upper bound function of the criterion;
A convergence determination unit that repeats the update by the parameter update unit until a predetermined convergence condition is satisfied;
The signal analysis device according to claim 1, comprising: A signal analysis method in a signal analysis device including a time frequency expansion unit and a parameter estimation unit,
The time frequency expansion unit inputs time series data of an observation signal in which a speech 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 base spectrum representing a power spectrum at each base and frequency of a pre-learned speech signal, and a cepstrum space. Based on the parameter representing the probability distribution of the cepstrum feature of the voice signal
Observation time frequency components of each time and frequency, the base spectrum of the audio signal, and the time frequency component of each time and frequency obtained from an activation parameter representing the power of the audio signal at each time, and the noise signal The base spectrum of the speech signal based on the distance from the sum of the time frequency components of each time and each frequency obtained from the base spectrum and the activation parameter of the speech signal, and the probability distribution of the cepstrum feature amount of the speech signal The activation parameter of the audio signal so as to reduce the criterion expressed using the regularization term representing the likelihood of the cepstrum feature quantity of the time-frequency component of each time and each frequency obtained from the activation parameter; The basis spectrum 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 represented by the following expression.

Where U ^(s) represents the activation parameter of the audio signal, H ⁽ⁿ⁾ represents the base spectrum of the noise signal, and U ⁽ⁿ⁾ represents the activation parameter of the noise signal. Y is the observation spectrogram, X is the time frequency component X ^{(s) of} each frequency obtained from the base spectrum of the speech signal and the activation parameter U ^(s), and the noise signal Represents the sum of the time-frequency components of each frequency obtained from the base 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 cepstrum feature quantity of the time-frequency component X ^(s) of each frequency obtained from the base spectrum of the speech signal and the activation parameter U ^(s) , and λ is a regularization parameter. By the parameter estimation unit estimating,
A parameter updating unit updates the activation parameter of the speech signal, the base spectrum of the speech signal of the noise signal, and the activation parameter so as to reduce an auxiliary function that is an upper bound function of the criterion. ,
The signal analysis method according to claim 4, wherein the convergence determination unit includes repeating the update by the parameter update unit until a predetermined convergence condition is satisfied.   The program for functioning a computer as each part of the signal analyzer of any one of Claims 1-3.