JP2001067094A - Voice recognizing device and its method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognizing device and its method capable of reducing deterioration in recognition performance due to a change in distance between an input terminal of a voice signal and a noise source, and due to variations in environmental noise. SOLUTION: This voice recognizing device is equipped with a spectrum computing means 101 for obtaining a noise-superimposed voice spectrum time series, an average spectrum computing means 102 for obtaining a noise spectrum by estimating a spectrum of superimposed noise from non-vocal zones, a noiseremoved spectrum group computing means 201 for obtaining noise-removed vocal-spectrum time series of a plurality of kinds by changing a scaling factor relative to the noise spectrum, a characteristic vector group computing means 202 for converting the noise-removed vocal spectrum time series of two or more kinds into characteristic vector time series of two or more kinds, a collation model memory 205 for memorizing a noiseless voice pattern and a model representing transition of the kinds of the characteristic vectors, and a three-dimensional collation means 203 for collating the noiseless voice pattern with the model representing the transition of the kinds of the characteristic vectors in a three-dimensional space made up of three axes, time, state, and characteristic vector.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus and method for speech uttered in a noisy environment and overlaid with noise.

[0002]

2. Description of the Related Art Background noise is superimposed on speech uttered in a noise environment, and the speech recognition rate is degraded. As a simple and effective method for removing the superimposed noise, a spectral subtraction method is widely used. Here, as an example, a description is given of a conventional speech recognition apparatus using the spectral subtraction method described in the document “Acoustic Engineering Course edited by the Acoustical Society of Japan, 7 Revised Speech” (Kazuo Nakada, Corona, pp. 130-131). I do.

FIG. 8 is a block diagram showing a configuration of a conventional voice recognition device. In FIG. 8, reference numeral 101 denotes a spectrum calculating means for performing spectrum analysis on a noise-superimposed speech input and extracting and computing a noise-superimposed speech spectrum time series;
Is an average spectrum calculating means for averaging the spectrum of the non-voice section and outputting the result as a noise spectrum; 103, a noise removing spectrum calculating means for subtracting the noise spectrum from the noise-superimposed voice spectrum time series to output a noise removing spectrum time series; A feature vector calculating means for obtaining a feature vector time series from the removed spectrum time series; 105, a matching model memory for storing a noise-free speech pattern for matching; and 106, a noise stored in the matching model memory 105 for the feature vector time series. This is a matching unit that performs a matching process with a no-voice pattern and outputs a recognition result that gives the maximum likelihood.

[0004] The operation of the conventional speech recognition apparatus will be described below. The spectrum calculating means 101 calculates a power spectrum by Fourier transform at predetermined time intervals with respect to the input of the noise-added speech, and outputs the power spectrum as a time series of the noise-added speech spectrum. Also, average spectrum calculation means 1
In No. 02, a noise-superimposed speech spectrum of several frames extracted from a non-speech section in the noise-superimposed speech spectrum time series, for example, immediately before a speech section or a pause section during speech production, is averaged for each frequency, and the noise spectrum is calculated. Output as The noise removal spectrum calculation means 103 subtracts the noise spectrum from each noise-superimposed speech spectrum in the time series of the noise-superimposed speech spectrum.

Here, the relationship between the power S (ω) at the frequency ω of the noise-removed voice spectrum, the power X (ω) at the frequency ω of the noise-superimposed voice spectrum, and the power N (ω) at the frequency ω of the estimated noise spectrum is shown. This is as shown in equation (1).

[0006]

(Equation 1)

Note that α is a parameter called a subtract coefficient, which represents the degree of noise component removal, and is usually adjusted to maximize recognition accuracy. Also, max ｛｝
Is a function that returns the element with the largest value among the elements in parentheses.

[0008] The feature vector computing means 104 converts the noise-reduced speech spectrum time series output from the noise-removing spectrum computing means 103 into an LPC (Linear Predictive Codin).
g) Convert to a vector expressing acoustic features in speech recognition such as cepstrum.

[0009] The matching means 106 compares the feature vector time series output from the feature vector calculating means 104 with the noise-free voice pattern stored in the matching model memory 105 to recognize a recognition candidate giving the maximum likelihood. Output as result. Here, as an example of the matching means, the document “Basic of speech recognition (below)” (Lawrence Rabiner, Biing-
Hwang Juang, NTT Advanced Technology Corporation, p. 125-128), a method of calculating the maximum likelihood using a Viterbi search in a speech recognition apparatus using a hidden Markov model (hereinafter referred to as HMM) will be described.

That is, the feature vectors from time 1 to time T
Time series Y = (y₁, Y_Two, ..., y _T) For the likelihood
One large optimal state sequence q = (q₁, Q_Two, ...,
q_TViterbi search to find the following four steps
Consists of

STEP 1 (initialization)

[0012]

(Equation 2)

[0013]

(Equation 3)

STEP 2 (repeated)

[0015]

(Equation 4)

[0016]

(Equation 5)

STEP 3 (End)

[0018]

(Equation 6)

[0019]

(Equation 7)

STEP 4 (backtrack)

[0021]

(Equation 8)

Here, δ _t (i) is the maximum likelihood at time t on one path and is expressed by the following equation.

[0023]

(Equation 9)

In the equations (2) to (8), Ψ _t (j) is an array for storing the argument of the path that maximizes the equation (9) at each time t and each state j. Also, a _{ij changes} from state i to state j.
, B _i (y _t ) is the output probability of the feature vector y _t in state i, π _i is the probability of being in state i in the initial state, λ is the voice model for verification, and each is a noise-free environment. Learned from voice data uttered below.

In a general speech recognition apparatus, the state transition of a collation speech pattern is represented by a left-to-right HMM model with a restriction on the state transition as shown in FIG. Note that b _i (y) is the output probability of the feature vector y in the state i.

The state transition of the matching voice pattern is defined as Left-to-
Viterbi search when expressing with the right HMM model
This is shown in FIG. FIG. 10 shows the state at time t and state j.
Maximum likelihood δ_t(J) is at time t-1, state j
Maximum likelihood δ_t-1(J) at time t-1, state j-1
Maximum likelihood δ _t-1From (j-1), the likelihood is maximized.
Is calculated by selecting the appropriate path.
ing.

By the above operation, it is considered that the average spectrum of the noise section of the non-voice section is superimposed on the spectrum time series of the input noise-superimposed speech signal, and the noise component is removed from the power spectrum. A matching process with the no-matching model is performed to obtain a recognition result.

[0028]

Since the conventional under-noise speech recognition apparatus using the spectral subtraction method is configured as described above, it is superimposed on the average spectrum of noise immediately before utterance or the like and the actual speech section. When the noise spectrum difference is small, that is, when the fluctuation of the environmental noise is small, the operation is relatively good. However, when the noise source is a moving object and the distance from the input end of the audio signal to the noise source changes, or when the environmental noise is unsteady and has large fluctuations, the estimated noise spectrum and the noise are actually superimposed on the voice. There is a problem that an estimation error with respect to the noise spectrum increases and the recognition performance deteriorates.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and has as its object to obtain a speech recognition apparatus and method capable of reducing the degradation of recognition performance due to a change in the distance between a speech signal input terminal and a noise source. It is an object. It is another object of the present invention to provide a speech recognition apparatus and method capable of reducing recognition performance deterioration due to fluctuations in environmental noise.

[0030]

SUMMARY OF THE INVENTION A speech recognition apparatus according to the present invention is a speech recognition apparatus for analyzing a spectrum of a noise-superimposed input speech signal including a non-speech section to obtain spectrum characteristic parameters and performing speech recognition processing. Spectrum calculating means for analyzing the spectrum of the voice signal and outputting a noise-superimposed voice spectrum time series; and estimating the spectrum of the superimposed noise from the non-voice section in the noise-superimposed voice spectrum time series output from the spectrum calculating means. Means for calculating the average of the noise spectrum output from the spectrum calculating means and the noise spectrum output from the average spectrum calculating means when subtracting the noise spectrum output from the noise spectrum from the time series. The noise-removed speech spectrum time series And the noise removal spectrum group arithmetic means,
A feature vector group calculating means for converting a plurality of types of noise-removed speech spectrum time series output from the noise removing spectrum group calculating means into a plurality of types of feature vector time series;
A collation model memory storing a model representing a transition of the type of a feature vector and a noise-free speech pattern learned using speech data uttered in a noise-free environment, and output from the feature vector group calculating means. For a plurality of types of noise-removed speech feature vector time series, in a three-dimensional space consisting of three axes of time, state, and feature vector type, the noise-free speech pattern and the feature vector type stored in the matching model memory are stored. And a three-dimensional matching means for comparing the model with the model representing the transition and outputting a recognition result.

The apparatus further comprises a noise spectrum memory for storing a noise spectrum output from the average spectrum calculating means and a plurality of types of noise spectrum patterns previously learned from a large amount of noise data by using a clustering technique. The spectrum calculating means calculates a plurality of types of magnifications for the noise vector and a plurality of types of noise spectrums stored in the noise spectrum memory from each of the noise-superposed speech spectra of the noise-superposed speech spectrum output from the spectrum calculating means. It is characterized in that a plurality of types of noise-removed speech spectra are obtained by combining with a pattern.

Further, the collation model memory is characterized in that a model which does not restrict the transition of the type of the feature vector is stored as a model representing the transition of the type of the feature vector.

The matching model memory is characterized by storing an elgotic hidden Markov model capable of transitioning to all types as a model that does not impose restrictions on the transition of the types of feature vectors.

The matching model memory is characterized by storing a model in which the transition of the type of the feature vector is restricted as a model representing the transition of the type of the feature vector.

Further, the matching model memory is characterized in that a hidden Markov model capable of transitioning only between adjacent types of feature vectors is stored as a model in which transitions of types of feature vectors are restricted. .

Further, according to the speech recognition method of the present invention, in the speech recognition method for analyzing the spectrum of a noise-superimposed input speech signal including a non-speech section to obtain a spectrum feature parameter and performing speech recognition processing, A spectrum calculation step of performing analysis to obtain a noise-superimposed speech spectrum time series, and an average spectrum calculation of estimating a spectrum of superimposed noise from a non-speech section in the noise-superimposed speech spectrum time series obtained in the spectrum calculation step to obtain a noise spectrum. And a plurality of types of noise-removed voice spectrum time series by changing the magnification for the noise spectrum when subtracting the noise spectrum obtained in the average spectrum calculation step from the noise superimposed voice spectrum time series obtained in the spectrum calculation step. Obtained noise removal spectrum group calculation process A feature vector group calculating step of converting a plurality of types of noise-removed voice spectrum time series obtained in the noise removing spectrum group calculating step into a plurality of types of feature vector time series; and a plurality of types of feature vector groups obtained in the feature vector group calculating step. For the noise-removed speech feature vector time series, three axes of time, state, and feature vector type
A three-dimensional matching step of comparing a noise-free speech pattern learned using speech data uttered in a noise-free environment in a three-dimensional space with a model representing a transition of the type of a feature vector to obtain a recognition result thereof; It is characterized by having.

In addition, the noise removal spectrum calculation step includes a plurality of types of magnifications with respect to the noise vector and a large amount of noise data in advance from each noise superimposed speech spectrum of the noise superimposed speech spectrum time series obtained in the spectrum calculation step. It is characterized in that a plurality of types of noise-removed speech spectra are obtained by combining a plurality of types of noise spectrum patterns learned using a clustering method.

Further, the three-dimensional matching step is characterized in that a model that does not restrict the transition of the type of the feature vector is used as a model representing the transition of the type of the feature vector.

Further, the three-dimensional matching step is characterized in that an elgotic hidden Markov model capable of transitioning to all types is used as a model that does not impose restrictions on the transition of the types of the feature vectors.

Further, the three-dimensional matching step is characterized in that a model in which the transition of the type of the feature vector is restricted is used as the model representing the transition of the type of the feature vector.

Further, the three-dimensional matching step is characterized in that a hidden Markov model capable of transitioning only between adjacent types of feature vectors is used as a model in which transitions of types of feature vectors are restricted. is there.

[0042]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 FIG. 1 is a block diagram showing a configuration for explaining a speech recognition apparatus and method according to Embodiment 1 of the present invention. In FIG. 1, FIG.
Are denoted by the same reference numerals, 101 is a spectrum calculating means for performing spectrum analysis on the noise-superimposed speech input and extracting a noise-superimposed speech spectrum time series, and 102 is the spectrum computing means. 101
Average spectrum calculating means for averaging the spectrum of a non-speech section in the noise-superimposed speech spectrum time series output from, and outputting as a noise spectrum.

As a new code, reference numeral 201 denotes the average spectrum calculating means 102 based on the noise-superimposed speech spectrum time series output from the spectrum calculating means 101.
The noise spectrum is subtracted by changing the magnification for the noise spectrum when subtracting the noise spectrum output from
A noise removal spectrum group calculating means for outputting a plurality of types of noise removal spectrum time series; a feature vector group calculating means for converting a plurality of types of noise removal spectrum time series into a plurality of types of feature vector time series; For a plurality of types of noise-removed speech feature vector time series output from the group calculation means 202, a collation model memory 205 described later stores in a three-dimensional space composed of three axes of time, state, and feature vector type. Three-dimensional matching means for matching a noise-free voice pattern with a model representing a transition of the type of a feature vector and outputting a recognition result; and 205, a noise-free voice trained using voice data generated in a noise-free environment This is a collation model memory that stores a model representing the transition between the types of patterns and feature vectors.

The speech recognition apparatus according to the first embodiment shown in FIG. 1 is constituted by the above-described block diagram shown in FIG. 1. The steps constituting the corresponding speech recognition method are as follows. Process. a. A spectrum calculation step of performing spectrum analysis on the noise-superimposed input speech to obtain a noise-superimposed speech spectrum time series; b. An average spectrum calculating step of estimating a spectrum of superimposed noise from a non-voice section in the noise-superimposed voice spectrum time series obtained in the spectrum calculating step and obtaining the spectrum as a noise spectrum; c. Noise removal to obtain a plurality of types of noise-removed speech spectrum time series by changing the magnification of the noise spectrum obtained when the noise spectrum obtained in the average spectrum calculation step is subtracted from the noise-superimposed speech spectrum time series obtained in the spectrum calculation step Spectrum group calculation step, d. A feature vector group calculating step of converting a plurality of types of noise-removed speech spectrum time series obtained in the noise removing spectrum group calculating step into a plurality of types of feature vector time series; e. A plurality of types of noise-removed speech feature vector time series obtained in the feature vector group calculation step were uttered in a three-dimensional space composed of three axes of time, state, and feature vector in a noise-free environment. Matching is performed between the noise-free speech pattern learned using the speech data and the model representing the transition of the type of the feature vector to obtain the recognition result 3
Dimension matching process.

Next, the operation of the first embodiment having the above configuration will be described. The operations of the spectrum calculation means 101 and the average spectrum calculation means 102 are the same as the operations of the conventional example, and the description is omitted here. The noise removal spectrum group calculating means 201 uses V (multiple types) subtraction coefficients α ^{(k )} and (1 ≦ k ≦ V) from each of the time-series noise-superimposed voice spectra of the noise-superimposed voice spectrum. The spectrum is subtracted to obtain V kinds of noise-removed speech spectra S ^(k) (ω). Here, the value of α ^(k) is set in 0.5 steps as follows.

[0046]

(Equation 10)

Here, S ^(k) (ω) is the power at the frequency ω of the k-th noise-removed voice spectrum, and X (ω)
Represents the power at the frequency ω of the noise-superimposed speech spectrum. In this way, V types of noise-removed speech spectrum time series S ⁽¹⁾ (ω), S ⁽²⁾ (ω),..., S ^(v) (ω) (where S ^(k) (ω) = (S ₁ ^(k) (ω), S
₂ ^(k) (ω),..., _ST ^(k) (ω))).

The feature vector group calculating means 202 outputs V kinds of noise-removed speech spectrum time series S ⁽¹⁾ (ω), S output by the noise removing spectrum group calculating means 201.
⁽²⁾ (ω),..., S ^(v) (ω) is changed to L
V type feature vector time series Y ⁽¹⁾ , Y ⁽²⁾ , representing acoustic features in speech recognition such as PC cepstrum
···, Y ^{(v) (where _{Y (k) = Y 1 (}} k), Y 2 (k), ··
, Y _T ^(k) ).

[0049] In three-dimensional matching process unit 203, V type of time-series feature vector Y output from the feature vector group arithmetic unit ^{202 (1), Y (2} ), ···, against the Y ^(v), the time ,
Matching is performed in a three-dimensional space including three axes of the state and the type of feature vector, and a recognition candidate that gives the maximum likelihood is output as a recognition result.

The transition of the type of the feature vector is shown in FIG.
Expressed by an elgotic HMM model. In FIG. 2, c _kl
Is the transition probability from the type k of the feature vector to the type 1 of the feature vector, and each state is connected by a null transition that does not output the observation event. The reason why the elgotic HMM model is used is that the first embodiment does not restrict the transition of the type of the feature vector.

A sequence (q, v) = (q) of one optimal combination of the type of the state and the feature vector having the maximum likelihood
₁ , v ₁ ), (q ₂ , v ₂ ),..., (Q _T , v _T )
Perform a Viterbi search extended to the dimension.

STEP 1 (initialization)

[0053]

[Equation 11]

[0054]

(Equation 12)

STEP 2 (Repeat)

[0056]

(Equation 13)

[0057]

[Equation 14]

STEP 3 (End)

[0059]

(Equation 15)

[0060]

(Equation 16)

STEP 4 (Back Track)

[0062]

[Equation 17]

Here, δ _t (i, k) is the time t, the state i, and the characteristic vector on one path in a three-dimensional space composed of three axes of time, state, and feature vector. This is the maximum likelihood of the type k, and is represented by the following equation.

[0064]

(Equation 18)

In equations (11) to (17), Ψ
_t (j, l) stores the argument of the path that maximizes the expression (18) at each time t, each state j, and the type 1 of the feature vector.
It is a dimensional array. B _i (y _t ^(k) ) is the output probability of feature vector y _t ^(k) in state i, c _kl is the transition probability from feature vector type k to feature vector type l, ρ _k
Is the probability that the type of the feature vector is k in the initial state.

FIG. 3 shows the state transition of the verification voice pattern as Le.
3 in the case of expressing with ft-to-right HMM model and expressing the transition of the type of feature vector with elgotic HMM model
This shows the state of the dimensional Viterbi search.

FIG. 4 is a timing chart of FIG.
FIG. 6 is a diagram in which a range of t is extracted, and the maximum likelihood δ _t (j, l) at time t, state j, and feature vector type 1 is the maximum likelihood at time t−1, state j, and feature vector type k. Degree δ _t−1 (j, k) (where (1 ≦ k ≦ V)) and the maximum likelihood δ _t−1 (j−1) at time t−1, state j−1, and feature vector type k. , K) (where (1 ≦ k ≦
V)), the calculation is performed by selecting a path that maximizes the likelihood.

The operation and effect of the first embodiment will be described below. In a conventional noisy speech recognition device, it is assumed that the noise spectrum estimated from the non-speech section is uniformly superimposed on the entire speech section, and the only one that is adjusted to maximize the recognition performance for the evaluation data The value of the subtraction coefficient α was used. However, when the distance between the noise source and the voice input terminal fluctuates with time, the power of the noise spectrum superimposed on the voice at a certain time is different from the power of the noise spectrum at the time of noise estimation. In some cases, the sound is over-pulled, and an accurate noise-free speech spectrum cannot be obtained. As a result, a mismatch with the noise-free speech pattern occurs, and the recognition rate deteriorates.

Document "Speech recognition under non-stationary noise by parallel HMM method and spectral subtraction"
(Ryuji Mine, IEICE Transactions (D-II), Vo
l. J-78-D-II, No. 7, pp. 1021-1
027, 1995), the noise HMM was transformed into an elgotic HMM.
And performs recognition processing on the noise-removed speech feature vector after spectral subtraction in a three-dimensional space of the time, the state of the speech model, and the state of the noise model, so that the recognition performance in an unsteady noise environment Has been improved. However, since the above document does not describe the value of the subtraction coefficient, and in the first embodiment, the transition of the type of the feature vector is modeled instead of the noise model. It can be said that there is.

In the speech recognition apparatus and method according to the first embodiment, V types of subtraction coefficients α ^(k) are provided at each time t.
There are V types of feature vector candidates calculated using Since the type k of the feature vector at each time t is selected such that the likelihood is maximized, the noise spectrum may be overdrawn or overdrawn even if the distance between the noise source and the voice input terminal fluctuates. Can be prevented, and the deterioration of the recognition rate can be suppressed.

Further, in the speech recognition apparatus and method according to the first embodiment, as a model representing the transition of the type of the feature vector, transition to all types is possible without restricting the transition of the type of the feature vector. Although the elgotic HMM model is used, the subtraction coefficient α ^(k) at the time of noise removal is used as a model that restricts the transition of the type of feature vector.
Can be appropriately modeled by using the HMM model shown in FIG. 5 in which the value of can be changed only between the types of adjacent feature vectors.

Embodiment 2 Next, FIG. 6 is a block diagram showing a configuration for explaining a speech recognition apparatus and method according to Embodiment 2 of the present invention. 6, the same parts as those in the first embodiment shown in FIG.
The description is omitted. As a new code, reference numeral 204 denotes a noise spectrum memory that stores a noise spectrum output from the average spectrum calculation unit 102 and a plurality of types of noise spectrum patterns previously learned from a large amount of noise data using a clustering method. The calculating means 201 includes a plurality of types of magnifications for noise vectors from each of the noise-superimposed speech spectrums of the noise-superimposed speech spectrum time series output from the spectrum computing means 101, and a plurality of kinds of noise spectrum patterns stored in the noise spectrum memory 204. In order to obtain a plurality of types of noise-removed speech spectra.

The speech recognition apparatus according to the second embodiment is constituted by the above-described block diagram shown in FIG. 6. The steps constituting the corresponding speech recognition method are the same as those of the above-described embodiment. The noise removal spectrum calculation step according to 1 is performed from each noise-added speech spectrum of the noise-added speech spectrum time series obtained in the spectrum calculation step.
The only difference is that a plurality of types of magnifications for the noise vector are combined with a plurality of types of noise spectrum patterns previously learned from a large amount of noise data by using a clustering method to obtain a plurality of types of noise-removed speech spectra.

Next, the operation of the second embodiment according to the above configuration will be described. The operations of the spectrum calculation means 101 and the average spectrum calculation means 102 are the same as the operations of the conventional example, and the description is omitted here. In noise spectrum memory 204, learned from the noise spectrum average spectrum calculating unit 102 outputs and advance a large amount of noise data by using clustering techniques, stores V ₂ kinds of representative noise spectrum pattern.

The noise-removed spectrum group calculating means 201 calculates V ₁ types of subtraction coefficients α ^(k1) ,
(1 ≦ k ₁ ≦ V ₁ ) and V ₂ kinds of noise spectrum patterns N _k2 (ω), (1 ≦ k ₂ ≦ V ₂ )
= V ₁ V _Two types of noise-removed speech spectra S ^(k) (ω),
(1 ≦ k ≦ V) is obtained. Here, as below,
Set the value of α ^{(k1) in} 5 steps.

[0076]

[Equation 19]

Here, S ^(k) (ω) is the power at the frequency ω of the k-th noise-removed voice spectrum, and X (ω)
Is the power at the frequency ω of the noise-superimposed speech spectrum,
N (ω) represents the power at the frequency ω of the estimated noise spectrum. In this way, V types of noise-removed speech spectrum time series S ⁽¹⁾ (ω), S ⁽²⁾ (ω),.
·, S ^(V) (ω) (where S ^(k) (ω) = (S
₁ ^(k) (ω), S ₂ ^(k) (ω),..., _ST ^(k) (ω)).

The operations of the feature vector group calculating means 202 and the three-dimensional collating means 203 are the same as those of the first embodiment, and the description is omitted here.

Hereinafter, effects of the speech recognition apparatus and method according to the second embodiment will be described. In the conventional noisy speech recognition device, it is assumed that a noise spectrum estimated from a non-speech section is uniformly superimposed on all speech sections. But,
When the spectrum pattern superimposed on the voice fluctuates with time, such as in a non-stationary noise environment such as in a running car, the noise spectrum pattern superimposed on the voice at a certain time becomes the noise spectrum Since it is different from the pattern, an accurate noise-removed speech spectrum cannot be obtained. As a result, a mismatch with a noise-free voice pattern occurs, and the recognition rate deteriorates.

Further, the speech recognition apparatus and method according to the first embodiment can cope with fluctuations in spectrum power, but use only a single noise spectrum pattern.
It cannot deal with fluctuations in the spectral pattern. In the speech recognition apparatus and method according to the second embodiment, each time t is calculated using V ₁ types of subtraction coefficients α ^(k1) and V ₂ types of noise spectrum patterns N _k2 (ω). V = V ₁ V There are _two types of feature vector candidates.
Since the type k of the feature vector at each time t is selected such that the likelihood is maximized, the recognition rate deteriorates even if the distance between the noise source and the voice input terminal or the noise spectrum pattern superimposed on the voice fluctuates. Can be suppressed.

Further, in the speech recognition apparatus and method according to the second embodiment, as a model representing the transition of the type of the feature vector, transition to all types is possible without restricting the transition of the type of the feature vector. Although the elgotic HMM model is used, the noise spectrum pattern N _k2 (ω) at the time of noise removal is similar or the subtract coefficient α ^(k By using the HMM model shown in FIG. 7 in which the value of ( ⁾⁾ can be changed only between the types of adjacent feature vectors, it is possible to appropriately model the temporal change of the noise spectrum and the temporal change of the superimposed noise power. It is.

[0082]

As described above, according to the present invention, there are a plurality of types of feature vector candidates calculated using a plurality of types of subtract coefficients at each time, and the type of feature vector at each time is Since the likelihood is selected to be the maximum, it is possible to prevent the noise spectrum from being pulled too much or too much even if the distance between the noise source and the voice input terminal fluctuates, and to suppress the deterioration of the recognition rate. As a result, it is possible to reduce recognition performance deterioration due to a change in the distance between the input end of the audio signal and the noise source.

Further, even if the noise spectrum pattern to be superimposed on the voice fluctuates, the deterioration of the recognition rate can be suppressed.
Recognition performance degradation due to fluctuations in environmental noise can be reduced.

Further, by using a model that does not limit the transition of the type of the feature vector as a model representing the transition of the type of the feature vector, the deterioration of the recognition rate can be suppressed.

As a model that does not limit the transition of the types of feature vectors, elgo that can transition to all types
By using the tic HMM model, it is possible to suppress the deterioration of the recognition rate.

By using a model in which the transition of the type of the feature vector is restricted as a model representing the transition of the type of the feature vector, the temporal change of the superimposed noise power can be appropriately modeled. .

Further, as a model in which the transition of the type of the feature vector is restricted, an HMM model in which transition is possible only between the types of adjacent feature vectors is used.
The temporal change of the superimposed noise power can be appropriately modeled.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration for describing a speech recognition device and method according to Embodiment 1 of the present invention.

FIG. 2 is a diagram for explaining the speech recognition apparatus and method according to the first embodiment of the present invention, and is an explanatory diagram of an elegant HMM model representing a transition of a type of a feature vector.

FIG. 3 is a diagram for explaining a speech recognition apparatus and method according to Embodiment 1 of the present invention, in which a state transition of a collation speech pattern is represented by a left-to-right HMM model, and a transition of a type of a feature vector is performed. FIG. 5 is an explanatory diagram showing a state of a three-dimensional Viterbi search when is represented by an elgotic HMM model.

FIG. 4 is an explanatory diagram in which a range from time t-1 to time t in FIG. 3 is extracted.

FIG. 5 is a diagram for explaining the speech recognition apparatus and method according to the first embodiment of the present invention, and is an explanatory diagram of an HMM model in which transition is possible only between types of adjacent feature vectors.

FIG. 6 is a block diagram showing a configuration for explaining a speech recognition apparatus and method according to Embodiment 2 of the present invention.

FIG. 7 is a diagram for explaining a speech recognition apparatus and method according to Embodiment 2 of the present invention, and is an explanatory diagram of an HMM model in which transition is possible only between types of adjacent feature vectors.

FIG. 8 is a block diagram showing a configuration of a conventional speech recognition apparatus.

FIG. 9 is an explanatory diagram expressing a state transition of a matching voice pattern in a conventional example by a left-to-right HMM model with a restriction on the state transition.

FIG. 10 shows the state transition of the voice pattern for verification as Left-to-
FIG. 9 is an explanatory diagram showing a state of a Viterbi search when expressed by a right HMM model.

[Explanation of symbols]

101 spectrum calculation means, 102 average spectrum calculation means, 201 noise removal spectrum group calculation means, 2
02 feature vector group calculating means, 203 three-dimensional matching means, 204 noise spectrum memory, 205 matching model memory.

Claims (12)

[Claims] 1. A speech recognition apparatus for performing spectrum analysis on a noise-superimposed input speech signal including a non-speech section to obtain a spectrum feature parameter and perform speech recognition processing. A spectrum calculating means for outputting; a spectrum calculating means for estimating a spectrum of superimposed noise from a non-voice section in a noise-superimposed voice spectrum time series output from the spectrum calculating means and outputting the spectrum as a noise spectrum; Noise removal spectrum which outputs a plurality of types of noise removal speech spectrum time series by changing the magnification of the noise spectrum when subtracting the noise spectrum output from the average spectrum calculation means from the noise superimposed speech spectrum time series output from Group operation means, and the noise A feature vector group calculating means for converting a plurality of types of noise-reduced speech spectrum time series output from the spectrum group calculating means into a plurality of feature vector time series, and learning using voice data uttered in a noise-free environment. A matching model memory storing a model representing the transition of the type of the feature vector and the noise-free speech pattern, and a plurality of types of noise-removed speech feature vector time series output from the feature vector group calculating means. In a three-dimensional space composed of three axes of time, state, and feature vector type, the noise-free speech pattern stored in the matching model memory is compared with a model representing the transition of the feature vector type, and the recognition result is obtained. A speech recognition device comprising: a three-dimensional collating unit that outputs the speech. 2. A speech recognition apparatus according to claim 1, wherein a noise spectrum output from said average spectrum calculation means and a plurality of types of noise spectrum patterns previously learned from a large amount of noise data using a clustering method are stored. A noise spectrum memory, wherein the noise elimination spectrum calculation means includes a plurality of types of magnifications for the noise vector from each of the noise superimposed speech spectra of the noise superimposed speech spectrum time series output from the spectrum calculation means; A speech recognition apparatus characterized in that a plurality of kinds of noise-reduced speech spectra are obtained by combining a plurality of kinds of noise spectrum patterns stored in a spectrum memory. 3. The speech recognition device according to claim 1, wherein the matching model memory stores a model that does not impose a restriction on the transition of the type of the feature vector as a model representing the transition of the type of the feature vector. A speech recognition device characterized by the following. 4. The speech recognition apparatus according to claim 3, wherein the matching model memory stores an elgotic hidden Markov model capable of transitioning to all types as a model that does not impose restrictions on the transition of the types of feature vectors. A speech recognition device characterized by the above-mentioned. 5. The speech recognition device according to claim 1, wherein the matching model memory stores a model in which the transition of the type of the feature vector is restricted as a model representing the transition of the type of the feature vector. A speech recognition device characterized by the following. 6. The speech recognition apparatus according to claim 5, wherein the matching model memory is a hidden Markov model capable of transitioning only between adjacent types of feature vectors, as a model in which transitions of types of feature vectors are restricted. A speech recognition device characterized by storing 7. A speech recognition method for performing a spectrum analysis on a noise-superimposed input speech signal including a non-speech section to obtain a spectrum feature parameter and perform speech recognition processing, wherein the spectrum analysis is performed on the noise-superimposed input speech and a noise-superimposed speech spectrum time series. A spectrum calculation step of estimating a spectrum of superimposed noise from a non-speech section in the noise-superimposed speech spectrum time series obtained in the spectrum calculation step, and obtaining an average spectrum as a noise spectrum. A noise removal spectrum group calculation step of obtaining a plurality of types of noise removal speech spectrum time series by changing the magnification for the noise spectrum when subtracting the noise spectrum obtained in the average spectrum calculation step from the noise superimposed speech spectrum time series, Above noise removal spectrum group calculation Vector operation sequence for converting a plurality of types of noise-removed speech spectrum time series obtained in the above process into a plurality of types of feature vector time series, and a plurality of types of noise-removed speech feature vector time series obtained in the feature vector group operation process On the other hand, in a three-dimensional space consisting of three axes of time, state, and type of feature vector, transition of a noise-free voice pattern and a type of feature vector learned using voice data uttered in a noise-free environment is performed. A three-dimensional matching step of performing matching with a represented model and obtaining a recognition result thereof. 8. The speech recognition method according to claim 7, wherein the noise-removed spectrum calculation step includes a step of calculating a plurality of noise-free speech spectrums of the noise-superimposed speech spectrum time series obtained in the spectrum calculation step. A speech recognition method characterized by obtaining a plurality of types of noise-removed speech spectra by combining a plurality of types of magnifications and a plurality of types of noise spectrum patterns previously learned from a large amount of noise data using a clustering method. 9. The speech recognition method according to claim 7, wherein the three-dimensional matching step includes, as a model representing the transition of the type of the feature vector, a model that does not restrict the transition of the type of the feature vector. A speech recognition method characterized by using: 10. The speech recognition method according to claim 9, wherein in the three-dimensional matching step, an elgotic hidden Markov model capable of transitioning to all types is used as a model that does not impose restrictions on the types of the feature vectors. A speech recognition method characterized by using: 11. The speech recognition method according to claim 7, wherein the three-dimensional matching step includes, as a model representing the transition of the type of the feature vector, a model in which the transition of the type of the feature vector is restricted. A speech recognition method characterized by using a character string. 12. The speech recognition method according to claim 11, wherein in the three-dimensional matching step, as a model in which a transition of a type of a feature vector is restricted, a hidden Markov that can transition only between adjacent types of a feature vector is provided. A speech recognition method characterized by using a model.