JPH09230886A - Noise-resistant hidden markov model creating method for speech recognition and speech recognition device using the method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make the multiplicative distortion of linear spectral area caused by passing through telephone line into an effective noise-resistance speech HMM(Hidden Markov Model). SOLUTION: Moises in an utterance environment are recorded (S1 ), and HMM of the noises a generated. Each output probability distribution of the noise HMM and a speech HMM created from a speech uninfluenced by the noise and the multiplicative distortion is transformed to a linear spectral area (S31 ), and the speech HMM distribution in the linear spectral area is multiplied by a multiplicative distortion W of an unknown. The result of this of this multiplication and the noise HMM distribution in the linear spectral area are convolution-operated (S322 ), and the operation result is inversely transformed (S33 ) to the original speech HMM area, and an incomplete noise-resistant speech HMM (S3 ) is generated with the multiplicative distortion mode as an unknown. Frequency to the input speech of this incomplete HMM is determined, to estimate (S4 ) the multiplicative distortion of the incomplete HMM to make the frequency maximum, and this estimated value is substituted for incomplete HMM to obtain the noise-resistant speech HMM (S5 ).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used in a voice recognition method using a Hidden Markov Model (hereinafter referred to as HMM), and is passed through voice added with background noise or through a telephone line, for example. Noise resistance H for voice recognition, which is suitable for recognizing voices with multiplicative distortion
The present invention relates to a method for creating an MM and a voice recognition device to which the method is applied.

[0002]

2. Description of the Related Art A conventional example of a voice recognition method using an HMM will be described with reference to FIG. For the voice input through the voice input unit 1, the voice recognition unit 2 calculates the degree of similarity with each voice HMM stored in the voice HMM storage unit 3 and recognizes the recognition result based on that value. It is output via the result output means 4.

Conventionally, the creation of a voice HMM is generally performed on the basis of voice information obtained in a noise-free state. Since the voice HMM obtained in this way is not affected by noise, it is distorted by passing through the input voice or telephone line affected by noise in a real environment where noise is present. When calculating the degree of similarity with the input voice, the voice HMM is not appropriate, and the voice recognition performance is significantly deteriorated.

On the other hand, it is also practiced to record a voice affected by noise in an environment where noise is present and create a voice HMM based on this voice, but the types of noise are enormous. In order to obtain high recognition performance, the entire structure of the voice recognition device becomes bloated. In addition, in a real environment, in order to record a learning voice and create a voice HMM from the recorded voice, the learning voice length needs to be long, for example, about 24 hours.
It takes about two months, for example, to create the MM. In this way, learning voices are recorded for each real environment,
Therefore, it is not possible to easily create a voice HMM in a short time.

From this point of view, in a real environment, a noise-resistant voice HMM suitable for the environment can be created in a short time and relatively easily, and a voice recognition that can obtain a high recognition rate even in a noisy environment. The method is document 1 (F.Matin, K.Sh
ikano & Y.Minami, “Recognition of Noisy Speech b
y Composition of Hidden Markov Models "ISSN1
018-4074 (Volume 2) ESCA, pp. 103
1-1034). Since the present invention is an improvement of this conventional method, a method of creating a noise resistant speech HMM used in this conventional method will be briefly described below.

According to this conventional method, noise is recorded in a real environment, a noise HMM is created based on this recording noise, and this noise HMM and a speech HMM not affected by noise or multiplicative distortion are combined in a product space. To do. This synthesis is performed as shown in FIG. Cepstral coefficients are widely used as acoustic parameters used in HMMs for speech recognition. The cepstrum coefficient has a relationship of logarithmic spectrum (logarithmic power spectrum) and cosine transform.

It is assumed that both the noise HMM and the voice HMM are created in the cepstrum domain. These noise HM
The distributions of the output probabilities of the M and the speech HMM are respectively cosine-transformed to calculate the distributions on the respective logarithmic spectra (S ₃₁₁ ). Next, the exponential transformation of these logarithmic spectrum distributions is performed to calculate the respective linear spectrum distributions ( _S312 ). The distribution of the noise HMM and the speech HMM, which are linear spectra, is obtained by a convolutional operation (S ₃₂ ), the composite HMM is logarithmically transformed (S ₃₃₁ ), and subsequently the inverse cosine transform is performed to perform noise resistance in the cepstrum domain. A voice HMM is created ( _S332 ).

The distribution of the output probabilities of the speech HMM and the noise HMM is a mixture of normal distributions (so-called Gaussiann mix).
ture). Since the normal distribution can be expressed by its mean value and covariance,
Each conversion in FIG. 2 performs conversion of average value and conversion of covariance. Next, the conversion method when the distribution of the output probability of the HMM is a normal distribution will be further described.

First, consider the 0th to pth cepstrum coefficients as parameters of the distribution of the output probabilities of these HMMs, and C = (C ₀ C ₁ C ₂ ... C _p-1 C _p ) (1) D = (D ₀ D ₁ D ₂ ... D _p-1 D _p ) ... (2) Here, C is the average value of the distribution of the voice HMM, and D is the average value of the distribution of the noise HMM. The conversion from this cepstrum coefficient to a logarithmic spectrum is known as a cosine transform, which is a linear transform and is (p + 1) × m.
Is represented by a conversion matrix (COS). m is the order of the log spectrum. When the vector of the logarithmic spectrum of the average value of each distribution of the voice HMM and the noise HMM is represented by LC and LD, respectively, LC = C (COS) ... (3) LD = D (COS) ... (4) (FIG. 2, S ₃₁₁ ). If the covariances of the speech HMM and noise HMM distributions in the cepstrum region are Σ ^C and Σ ^D , respectively, the covariances Σ ^LC and Σ ^LD of these distributions in the logarithmic spectrum region are Σ ^LC = (COS) Σ ^C (COS) ^t (5) Σ ^LD = (COS) Σ ^D (COS) ^t (6) ^t represents a transposed matrix. In this way, voice H
The average value and covariance of the logarithmic spectrum region of the normal distribution of the MM and the noise HMM are obtained (FIG. 2, S ₃₁₁ ).

Next, exponential conversion for converting a logarithmic spectrum into a linear spectrum will be described.
This transformation does not take the form of a normal distribution, but the transformed one is approximated using a normal distribution. Mean values LC and LD of the logarithmic spectrum region, mean values SC of respective distributions when respective covariances Σ ^LC and Σ ^LD are subjected to exponential transformation,
When SD and covariances Σ ^SC and Σ ^SD are calculated, SC ⁱ = exp (LC ⁱ + Σ ^LC _ij / 2) ... (7) SD ⁱ = exp (LD ⁱ + Σ ^LD _ij / 2) ... ( 8) Σ ^SC _ij = SC ⁱ × SC ^j × {exp (Σ ^LC _ij ) −1} (9) Σ ^SD _ij = SD ⁱ × SD ^j × {exp (Σ ^LD _ij ) −1} ... (10) i, j = 0,1,2, ..., p ( _S312 ).

Since ambient noise is additive noise and speech and noise can be added in the region of the linear spectrum, the speech HMM and noise HM in the region of the linear spectrum are present.
The average value M of the noise resistant speech HMM distribution, which is the sum of the distributions of M
And covariance Σ ^M are calculated by the following equation (FIG. 2, S ₃₂ ). M ⁱ = SC ⁱ + SD ⁱ (11) Σ ^M _ij = Σ ^SC _ij + Σ ^SD _ij (12) The average value M ⁱ and covariance Σ ^M of the distribution thus obtained have been calculated up to now. Reverse the process of to convert to the cepstrum region. First, logarithmic transformation, which is the inverse transformation of exponential transformation, is performed. Letting LM be the logarithmically transformed average value and Σ ^LM be the covariance, it is the inverse transformation of the exponential transformation, so LM ⁱ = log (M ⁱ ) −½ log (Σ ^M _ij / M ^{i 2} +1). ) ··· (13) Σ ^LM _ij = log (Σ ^M _ij / (M ⁱ M ^j ) +1) ··· (14) is calculated (FIG. 2, S ₃₃₁ ). Further, the inverse cosine transform (COS ′) m × (p + 1) is used to transform the logarithmic spectrum into the cepstrum region, and the average value S and covariance Σ ^S of the output probability distribution of the noise-resistant speech HMM are obtained by the following equation (FIG. 2). , S ₃₃₂ ).

S = LM (COS ′) (15) Σ ^S = (COS ′) Σ ^LM (COS ′) ^t (16) When the distribution is a single normal distribution, two distributions Then, the above conversion may be performed. When the distribution is a mixture of normal distributions, the above conversion may be performed for all combinations of distributions. Therefore, for example, when the speech HMM is a mixture of three normal distributions and the noise HMM is a mixture of three normal distributions, the noise-resistant speech HMM is a mixture of 3 × 3 = 9 normal distributions.

In the above description, the respective distribution forms in the HMM of speech and noise are taken up and the synthesis method thereof is described. The normal voice HMM can be represented by a model of three states A, B, and C that transit from right to left as shown in the upper left of FIG. On the other hand, as the noise model, an ergodic HMM that transits between two states 1 and 2 as shown in the upper right of FIG. 3 is suitable. At this time, the noise resistant speech HMM becomes a product model as shown in the lower part of FIG. 3, and has six states 1A, 1B, 1C, 2A, 2B and 2C, and each state is a state of the speech HMM and a noise HMM.
It consists of a combination of states. For example, the state 1A is synthesized from the state A of the voice HMM and the state 1 of the noise HMM, and the output distribution in this state 1A is P _A * P ₁ . Here, P _A is the distribution in the state A of the speech HMM, P ₁ is the noise H
Each of the distributions of MM in the state 1 is shown, and * is shown in FIG.
Represents the transformation described with reference to. Such an operation is performed using the distributions in the states of the voice HMM and the noise HMM corresponding to each of the states 1B, 1C, 2A to 2C. Further, the transition probability between the states of the noise resistant speech HMM is in the form of the product of the transition between the states of the speech HMM and the transition between the states of the noise HMM as shown in the lower part of FIG. For example,
The transition probability from the state 1A to the state 1B is a product of the transition probability a _AB from the state A of the speech HMM to the state B and the transition probability a ₁₁ of the noise HMM from the state 1 to the state 1 (a _AB × a ₁₁ )become.

According to this conventional method, it is possible to obtain a speech HMM which is somewhat strong against the additive noise. Moreover, voice H
An existing one can be used as a set of MMs, while a noise HMM can record a short time noise of, for example, 5 to 6 seconds, and at most about 20 seconds in a real environment, and can be created based on the recorded noise. Moreover, the noise HMM can be created in about 1 second. In this way, the environmental noise is recorded, and a noise HMM is further created.
MM and MM are combined as described above to create a noise resistant speech HMM, and the time until the input speech is recognized using this noise resistant speech HMM is about one minute (it is much shorter if a high-speed arithmetic means is used). It's a short time.

[0015]

However, this conventional method has no effect on the multiplicative distortion in the linear spectral region which occurs in a voice signal when the voice signal is passed through, for example, a telephone line, and the additive noise is not generated. It wasn't enough for me. It is an object of the present invention to obtain a high recognition rate not only for additive noise but also for voices that are subject to multiplicative distortion in the linear spectral region, and that can be produced relatively easily in a short time. It is to provide a method for creating a noise-resistant hidden Markov model for recognition.

Another object of the present invention is to obtain a high recognition rate not only for additive noise but also for speech that has undergone multiplicative distortion in the linear spectrum region, and for making noise resistant speech HMM relatively simple and in a short time. The object is to provide a voice recognition device that can be manufactured in

[0017]

According to the method of the present invention, a speech HMM that is not affected by noise or multiplicative distortion (hereinafter simply referred to as noise) and its H generated from noise.
From the MM (noise HMM), an incomplete noise-resistant speech HMM including a multiplicative distortion in the linear spectrum region or S / N (signal / noise) as an unknown (variable) is created in the first step, and the uncompleted noise-resistant speech HMM is generated. The multiplicative distortion or S / N that maximizes the likelihood of the input speech of
The noise-resistant voice HMM is completed in the third step by estimating the value in the step and substituting the estimated value into the uncompleted noise-resistant voice HMM.

The first step is the speech HMM and the noise HM.
The uncompleted noise resistant speech HMM is obtained by synthesizing M and M in the product space. In the synthesis in the product space, the distributions of the output probabilities of the speech HMM and the noise HMM are transformed into a linear spectral domain in the fourth step, and the distributions of the output probabilities of the speech HMM and the noise HMM in the linear spectral domain are convoluted with each other. The calculation is performed in the fifth step, and the convolution calculation result is inversely transformed into the original speech HMM area in the sixth step.

The distributions of the output probabilities of the speech HMM and the noise HMM respectively expressed in the cepstrum domain are cosine-transformed, and further exponential-transformed. In the sixth step, the convolution operation result is logarithmically transformed, Further, inverse cosine transform is performed. In the fourth step, the output probability distributions of the voice HMM and the noise HMM respectively expressed in the logarithmic spectrum domain are subjected to exponential transformation, and the sixth step is subjected to logarithmic transformation of the convolution operation result.

The synthesis in the product space is performed by the convolution operation of the distributions of the output probabilities of the speech HMM and the noise HMM respectively expressed in the linear spectral domain. In the second step, various multiplication distortions or S /
N is given, the likelihood between each of these uncompleted noise-resistant speech HMMs and the input speech is calculated, and the multiplication distortion or S / N that maximizes the calculated likelihood is used as the estimated value.

In the second step, estimation is performed by iterative calculation according to the maximum likelihood estimation method or the steepest descent method. The noise HMM used in the first step is created by recording environmental noise. The convolution operation of the distribution of the speech HMM and the distribution of the noise HMM in the linear spectrum region is performed by multiplying the linear spectrum of the distribution of the speech HMM by an unknown multiplicative distortion.

The convolution operation of the distribution of the speech HMM and the distribution of the noise HMM in the linear spectrum domain is performed by the speech HM.
Multiply the linear spectrum of the distribution of M by 10 ^{− (S / N) / 2} or 10 to the linear spectrum of the distribution of the noise HMM.
^It is performed by multiplying ^{(S / N) / 2} (S / N is an unknown number). An unknown multiplicative distortion component is added to each cepstrum of the output probability distribution of the speech HMM represented in the cepstrum region, and synthesis in the product space is performed.

The voice H represented by the cepstrum region
The 0th-order unknown S / N component of the cepstrum of the distribution of the output probabilities of one of the MM and the noise HMM is added to perform synthesis in the product space. An unknown multiplicative distortion component is added to the logarithmic spectrum of the output probability distribution of the speech HMM expressed in the logarithmic spectrum domain, and synthesis in the product space is performed.

The unknown S / N component is added to the logarithmic spectrum of the output probability distribution of one of the speech HMM and the noise HMM expressed in the logarithmic spectrum domain to perform synthesis in the product space. According to the speech recognition apparatus of the present invention, there is provided an apparatus for recognizing an input speech by using a noise resistant speech HMM made up of a speech HMM and a noise HMM by the method of the present invention, which is affected by noise and multiplicative distortion. A set of voice HMMs that have not been received are stored in the voice HMM storage unit, ambient noise is input by the noise input means, and noise HMMs are created by the noise HMM creating means based on the input noises and stored in the noise HMMs. Noise HMM stored in the storage unit and the voice HMM stored in the voice HMM storage unit
And the S / N or the multiplicative distortion from the S / N or the multiplicative distortion storage unit are processed by the incomplete noise resistant speech HMM synthesizing means.
The unfinished noise resistant HMM is created by synthesizing in the product space with S / N or the multiplicative distortion as an unknown number, and the unknown number of this unfinished noise resistant speech HMM is output from the voice input means by the S / N or the multiplicative distortion estimation means. The uncompleted noise-resistant speech HMM is estimated so that the likelihood of the uncompleted noise-resistant speech HMM for the input speech is maximized, and the estimated S / N or the value of the multiplicative distortion is converted into the uncompleted noise-resistant speech HMM by the noise-resistant speech HMM completion means. Noise-resistant speech HMM is completed by substitution and noise-resistant speech HMM
The input voice is stored in the storage unit, the degree of similarity with the noise resistant voice HMM is calculated by the voice recognition means, and the recognition result is output based on the calculation result.

[0025]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 4 shows an example of the functional configuration of a speech recognition apparatus according to the present invention, and FIG. 5A shows the processing procedure in an example of the present invention method. The voice HMM storage unit 3 stores a set of voice HMMs created for voice information recorded in advance in a noise-free state. And
Environmental noise is recorded via the noise input means 5 (S ₁ ), and noise HMM is created by the noise HMM creating means 6 for the recorded environmental noise (S ₂ ). The noise HMM can be created by the same method as that of the speech HMM, and this can be done by, for example, SELevinson, L. &. Rabinen, and MMSondhi.
: “An Introduction to the Application of the The
ory of Probabilistic Functions of a Markov Process
to Automatic Speech Recognition ”, The Bell Syste
m Technical Journal, Vol. 62, No.4 (April 1983)
Is shown in The created noise HMM is stored in the noise HMM storage unit 7. The noise HMM stored in the noise HMM storage unit 7 is represented by the same characteristic parameter as that of the speech HMM stored in the speech HMM storage unit 3, that is, noise is generated when the speech HMM is represented in, for example, a cepstrum region. The HMM is also expressed in the cepstrum region.

Incomplete noise resistant speech HMM synthesizer 8 generates noise H
From the noise HMM in the MM storage unit 7 and the voice HMM in the voice HMM storage unit 3, the S / N or multiplicative distortion storage unit 9 is obtained.
An uncompleted noise-resistant speech HMM including the multiplicative distortion or S / N in the linear spectral region in as an unknown is created (S ₃ ). This incomplete noise resistant speech HMM is created by noise HM
This is performed by synthesizing M and the voice HMM in the product space. At that time, the voice HMM is multiplied or added with the unknown multiplicative distortion, or the voice HMM or the noise HMM is multiplied or added with the unknown S / N. .

In the synthesis of the product space of the noise HMM and the speech HMM, for example, as in the conventional method, as shown in FIG. 5B, the distribution of the output probabilities of the noise HMM and the speech HMM is converted into a linear spectral domain (S ₃₁ ), The convolution operation of the distribution of the output probabilities of the noise HMM and the speech HMM in the linear spectrum region is performed. At this time, the multiplicative distortion (unknown number) is applied to the linear spectrum of the output probability distribution of the speech HMM in the linear spectrum region. Or a component corresponding to S / N in the linear spectrum of the output probability distribution of one of the speech HMM and the noise HMM in the linear spectrum domain (unknown)
(S ₃₂₁ ), and then, a convolution operation is performed between the output probability distribution that has been multiplied and the output probability distribution that is not multiplied (S ₃₂₂ ), and the result of the convolution operation is inversely transformed into the original area to perform the uncompleted noise immunity. get a voice HMM (S _33). Similar to the description of the conventional method shown in FIG. 2, it is assumed that both the voice HMM and the noise HMM are created in the cepstrum region. In FIG. 6, parts corresponding to those in FIG. A specific example is shown. That is, the average values C and D of the output probability distributions of the speech HMM and the noise HMM are represented by the equations (1) and (2), and these are cosine transformed to obtain the logarithmic spectrum L.
C and LD respectively, and the covariance of each distribution is Σ
^LC , Σ ^LD (S ₃₁₁ ), and these are respectively averaged values SC ⁱ , SD ⁱ , Σ ^SC _ij in the linear spectral region,
It is converted into Σ ^SD _ij (S ₃₁₂ ). In the general communication field, for example, an output X when a voice S uttered in the presence of environmental noise N is received by a microphone and passed through a transmission path is X =, where W is a distortion received by the microphone and the transmission path. WS
It is known to be represented as + N. In consideration of this point, in this embodiment, the distribution of the output probability of the voice HMM is multiplied by the unknown multiplicative distortion W in the multiplicative distortion storage unit 9 (FIG. 4) (S ₃₂₁ ), and this multiplied distribution and noise are multiplied. The convolution operation with the HMM is performed ( _S322 ). That is, equations (11) and (12)
Instead of, the following equation is calculated.

M ⁱ = W _i SC ⁱ + SD ⁱ (17) Σ ^M _ij = W _i W _j Σ ^SC _ij + Σ ^SD _ij (18) Results of this convolution operation M ⁱ , Σ ^M _ij For Figure 2
As in the case of, the logarithmic transformation is performed by the equations (13) and (14) (S ₃₃₁ ), and the inverse cosine transformation is performed by the equations (15) and (16) to calculate the multiplicative distortion in the original cepstrum region by an unknown number. An uncompleted noise resistant speech HMM including
₂₃₂ ).

Next, the multiplicative distortion W (= W ₁ , ..., Which is an unknown number in the incomplete noise-resistant speech HMM obtained in this way.
.., W _m ) is estimated (FIG. 5A, S ₄ ). Therefore, a voice is input from the voice input means 1 (FIG. 4), and the S / N or multiplicative distortion estimation means 10 calculates the likelihood of the set M (W) of uncompleted noise-resistant voice HMMs with respect to the input voice sequence X. Degree P
Estimate the multiplication distortion W that maximizes (X | M (W)). This estimation can be obtained by iterative calculation by the steepest descent method or the maximum likelihood estimation method.

That is, the likelihood P (X | M
(W)) is maximized by executing the following iterative calculation. 1. Initialize W, W ^t = W ^t-1 + ε (∂P (X | M (W)) / (∂
2.) estimate the next W ^t using Update W ^t-1 with W ^t .

Repeat steps 4.2 and 3 until convergence. Use an appropriate small value for ε. Likelihood P by the maximum likelihood estimation method
To maximize (X | M (W)), do as follows. Generally, the following equation holds between the trellis likelihood and the Viterbi likelihood. P (X | M (W)) = ΣP (X, S | M (W)) Σ is the sum for all S, where S represents a state transition. Further, Q (W, W ') is defined as follows, where W after updating W is W'.

Q (W, W ') = ΣP (X, S | M
(W)) log P (X, S | M (W ')) Σ is the sum of all S Here, if Q (W, W') > Q (W, W), then P (X | M
(W ')) > P (X | M (W)). An estimation method of W using this principle and the maximum likelihood estimation method will be shown below.

1. Initialize W, Q the ^{(W t-1, W t} ) estimated by maximum likelihood estimation method W ^t to maximum. 3. Update W ^t-1 with W ^t . Repeat steps 4.2 and 3 until convergence. The value of the multiplicative distortion W estimated in this way is calculated by the HMM.
The completion means 11 (FIG. 4) substitutes each incomplete noise resistant speech HMM to complete the noise resistant speech HMM, and this is completed.
It is stored in the M storage unit 12 (FIG. 5A, S ₅ ). The unknown voice is recognized by inputting the unknown voice from the voice input unit 1, calculating the degree of similarity with each noise resistant voice HMM stored in the noise resistant voice HMM storage unit 12 by the voice recognition unit 2, and The recognition result is output based on the calculation result. The input speech X used when estimating the multiplicative distortion is not a learning speech,
It may be the unknown voice that you are trying to recognize. In the latter case, the multiplicative distortion is estimated, and then the voice recognition process is performed on the unknown voice.

In the above specific example, a voice HMM and a noise HMM
Is the cepstrum region, but the logarithmic spectrum
This is also true for those made in the (logarithmic power spectrum) domain.
The invention can be applied. In this case, it's shown in Figure 7.
Sea urchin in step S in FIG.₃₁₂Exponency in
Distribution of noise HMM in logarithmic spectral domain
And the speech HMM distribution,
Obtain each distribution of the region (S₃₁), The linear distribution of the speech HMM
Multiply the spectrum by the multiplicative distortion W
(S ₃₂₁), The distribution of this multiplied speech HMM and the noise H
Perform the convolution operation with the distribution of MM (S₃₂₂),
The calculation result is logarithmically converted to obtain an incomplete noise resistant speech HMM.
(S₃₃).

Similarly, when the speech HMM and the noise HMM are created in the linear spectral domain, the uncompleted noise-resistant speech HMM is immediately obtained by performing convolution between these distributions. Furthermore, in the above, the multiplicative distortion W is introduced as an unknown into the incomplete noise resistant speech HMM, but S / N may be introduced. That is, the average value M ⁱ and the covariance Σ ^M _ij of the distribution of the incomplete noise-resistant speech HMM obtained as a result of the convolution of the distribution of the speech HMM and the distribution of the noise HMM in this case.
On the other hand, S / N is introduced as in the following equations.

M ⁱ = SC ⁱ + kSD (19) Σ ^M _ij = Σ ^SC _ij + k ² Σ ^SD _ij (20) k = 10 ^{− ((S / N) / 2)}・ ・ ・ ( 21) In this case, W is stored in the S / N or multiplicative distortion storage section 9 in FIG.
Instead, k is stored, and the multiplication in each step S ₃₂₁ of FIGS. 5B, 6 and 7 multiplies the distribution of the noise HMM (or the distribution of the speech HMM) by k.
That is, the equations (19), (20), and (21) are calculated. Further, in the S / N or multiplicative distortion estimating means 10 in FIG. 4, P (X
S / N is chosen to maximize the likelihood of | M (S / N)). That is, with respect to the input speech sequence X, the S / N that maximizes the likelihood of the set M (S / N) of the incomplete noise-resistant speech HMMs, which is a function of the S / N, is estimated. As the input voice, learning voice or recognition target voice uttered in the environment when the noise HMM is created may be used. This S / N estimation can also be obtained by iterative calculation by the maximum likelihood estimation method or the steepest descent method for the S / N that maximizes P (X | M (S / N)). Another example of S / N estimation is shown in FIG. N
Prepare individual S / N values (S / N) ₁ to (S / N) _N ,
Substituting these into equations (19), (20), and (21), N uncompleted noise-resistant speech HMMs are created, and the likelihood P (X | M ( S / N))
Respectively, and the maximum S / N among these N likelihoods is calculated.
Choose

As (S / N) ₁ to (S / N) _N , for example, in a noise recording environment for creating a noise HMM, the ratio of the sum of the signal S and the noise N to the noise N (S + N).
/ N is obtained, and three values of this value and each value obtained by ± 3 dB are used, and the likelihood P (X | M (S
/ N)), and the maximum S / N is determined. In addition, a formulation that maximizes the likelihood P (X, S | M (S / N)) by the Viterbi algorithm instead of P (X | M (S / N)) is also possible. Here, S represents the state transition of the HMM. The estimation of the multiplicative distortion W may also be performed as shown in FIG.

In the above, the introduction of the multiplicative distortion W and S / N of the incomplete noise resistant speech HMM was performed in the linear spectrum domain, but in the cepstrum domain, W is added to the equation (1), that is, C + W. = (C ₀ + W ₀ , C ₁ +
W ₁ , ... C _p + W _p ) may be obtained, and the speech HMM to which this W is added and the noise HMM may be combined in the product space. Also, C + k = (C ₀ + k, C ₁ , ..., C _p ),
(K = αlog (10 ^{(S / N) / 2} ), α is a constant determined by cosine transformation)
The M and the noise HMM may be combined in the product space, or k may be added to the noise HMM and the speech HMM in which the k is not added may be combined in the product space. Furthermore, W or S / N may be introduced in the logarithmic spectral domain. That is, LC + W = C (COS) + W is obtained,
Synthesis may be performed in the product space of this and LD, or LC + k = (LC ₀ + k, LC ₁ + k, ..., LC _P +
k), (k = log (10 ^{(S / N) / 2} )) and L
D may be combined in the product space with D, or LC
Alternatively, + k may be obtained and the product of LD and LD may be combined in the product space.

Further, in the above description, the noise HMM is the speech H.
I created one in the same area as MM.
May be obtained in the cepstrum region, and the noise HMM may be obtained in the logarithmic spectrum region. In this case, the amount of calculation is reduced by the calculation for converting the noise HMM from the cepstrum domain to the logarithmic spectrum domain.

[0040]

As described above, according to the present invention,
A noise HMM is created based on the noise of the utterance location, and S / N or multiplicative distortion is introduced as an unknown number from this noise HMM and a voice HMM created based on voice information recorded in advance in the absence of noise. The uncompleted noise resistant speech HMM is created, and the S / N or multiplicative distortion that maximizes the likelihood of the uncompleted noise resistant speech HMM with respect to the input speech is determined to create the noise resistant speech HMM. By doing this, it is possible to create a voice HMM that is resistant to S / N fluctuation, microphone distortion, line distortion, and speaker utterance fluctuation.
That is, the voice recognition device of the present invention using the voice HMM obtained by the method of the present invention can recognize a voice to which noise is added, a voice subjected to line distortion, and the like with a higher recognition rate than before.

Then, the multiplicative distortion is included in the speech HMM as an unknown value, and the synthesis of this and the noise HMM is performed in the product space, or S / N is included as an unknown value in one of the speech HMM and the noise HMM. This and
An uncompleted noise-resistant speech HMM that models noise-containing speech can be constructed by performing synthesis in the product space with an HMM that was not included as an unknown number, and the likelihood of this HMM for input speech is maximized. Noise resistant speech HM by determining multiplicative distortion or S / N
M can be configured.

Further, both the speech HMM and the noise HMM are converted into a linear spectrum region, and the distribution of the output probability of the speech HMM in the linear spectrum region is subjected to multiplicative distortion. The uncompleted noise resistant speech HMM is obtained by performing a convolution operation and inversely converting the result into the area of the original speech HMM. Similarly, both the voice HMM and the noise HMM are converted into a linear spectrum region, and the voice HMM and the noise HM in the linear spectrum region are converted.
For a linear spectrum of one output probability distribution of M, S
/ N component is multiplied, and the convolution operation of this and the output probability distribution of the HMM which is not multiplied by the S / N component is performed, and the calculation result is inversely transformed into the region of the original speech HMM to obtain the uncompleted noise resistant speech. HMM is required.

Speech HMM represented in the cepstrum domain
And the distribution of the output probability of the noise HMM are transformed to the linear spectral domain by the cosine transform and the exponential transform, then the convolution operation is performed, and the result of this operation is subjected to the inverse transformation, the logarithmic transformation and the inverse cosine transformation. , It is possible to calculate the distribution of uncompleted noise resistant speech HMM in the cepstrum region. Further, particularly, the output probability distributions of the speech HMM and the noise HMM expressed in the logarithmic spectrum domain are subjected to the exponential transformation, the convolution operation is performed, and the operation result is logarithmically transformed to obtain the incomplete noise resistance in the logarithmic spectrum domain. Voice HM
Find M.

A phoneme recognition experiment was conducted to investigate the effect of the method and apparatus of the present invention. As the evaluation data, 51 sentences of the telephone number guidance task uttered by one speaker were used. Noises of 12 dB and 6 dB were added to the experiment, and a voice distorted by multiplicative noise was used. The experimental results of the phoneme recognition rate are shown in the table below. Both are models of the cepstrum domain, and in the experiment of "speech HMM", the speech HM without any processing is performed.
M, that is, the voice HMM in the voice HMM storage unit 3 in FIG.
Was used. The "HMM synthesis method only" experiment is an experiment using the HMM obtained by the noise-resistant HMM creation method proposed in the above-mentioned Document 1 described in the background of the invention. “HMM synthesis + S / N estimation” is an experiment using an HMM created by introducing S / N as an unknown by the method of the present invention,
"HMM synthesis + multiplicative distortion estimation" is an HMM created by introducing the multiplicative distortion as an unknown by the method of the present invention.
It is an experiment using. HMMs in these inventive methods
The estimation of S / N and multiplicative distortion in the preparation of was performed by the steepest descent method. From these experimental results, the recognition rate of the conventional HMM synthesis alone is better than that of the speech HMM that does not process anything. According to the method of the present invention, it becomes stronger against additive noise, and moreover, against multiplicative distortion. It can be seen that also a noise resistant speech HMM that is significantly stronger than the conventional one is obtained and the recognition result is improved. As described above, according to the present invention, the noise model of the utterance environment is used and synthesized with the conventional speech HMM to produce an incomplete noise-resistant speech HMM in which multiplicative distortion or S / N is an unknown number, and the HMM is used. Robust HM suitable for the utterance environment because the maximum likelihood of the input speech is estimated
M can be created, and the speech recognition apparatus of the present invention using this HMM can achieve a high speech recognition rate. Moreover, according to the device of the present invention, noise H
The time from the creation of the MM to the recognition of the input voice is, for example, about 1 minute (which is shorter if a high-speed arithmetic device is used), and the recognition can be performed in a short time and with a relatively simple process.

[Brief description of drawings]

FIG. 1 is a block diagram showing a simplified functional configuration of a voice recognition device.

FIG. 2 is a flowchart showing a procedure of a conventional noise-resistant speech HMM creation process.

FIG. 3 is a diagram showing state transitions of a speech HMM, a noise HMM, and a noise-resistant speech HMM obtained by combining these.

FIG. 4 is a block diagram showing a functional configuration of an embodiment of a voice recognition device according to the present invention.

5A is a flowchart showing a processing procedure of an embodiment of the noise resistant speech HMM creating method according to the present invention, and B is a flowchart showing a procedure showing a concrete example of the processing of step S _{3 in} A. FIG.

FIG. 6 is a flowchart showing a further specific example of the processing in FIG. 5B.

FIG. 7 is a flowchart showing yet another specific example of the processing of FIG. 5B.

FIG. 8 is a diagram showing an example of a method for estimating the S / N of an incomplete noise-resistant speech HMM.

Claims (21)

[Claims] 1. A multiplicative distortion in a linear spectrum region is calculated from a speech HMM that is not affected by noise and multiplicative distortion (hereinafter simply referred to as noise) and its HMM (noise HMM) created from noise. A first step of creating an incomplete noise-resistant speech HMM including S / N (signal / noise) as an unknown number, and the above-mentioned multiplicative distortion or S / so that the likelihood of the incomplete noise-resistant speech HMM with respect to an input speech becomes maximum. A method for creating a noise-resistant hidden Markov model for speech recognition, comprising: a second step of estimating N; and a third step of substituting the estimated value into the incomplete noise-resistant speech HMM to complete the noise-resistant speech HMM. 2. The method of creating a hidden Markov model according to claim 1, wherein the first step includes a fourth step of giving the multiplicative distortion to the speech HMM, a speech HMM to which the multiplicative distortion has been given, and the noise HMM. And the fifth step of synthesizing and in the product space. 3. The hidden Markov model creation method according to claim 2, wherein the fourth step is addition of the multiplicative distortion to the average value of the distribution of the output probabilities of the speech HMMs represented in the cepstrum region. 4. The method of creating a hidden Markov model according to claim 2, wherein the fourth step is the addition of the multiplicative distortion to the average value of the output probability distribution of the speech HMM represented in the logarithmic spectrum domain. 5. The hidden Markov model creation method according to claim 1, wherein the first step comprises the speech HMM and the noise HMM.
To the HMM to which the component including the S / N is added and the HMM to which the component including the S / N is not added in the product space. 6. The method of creating a hidden Markov model according to claim 5, wherein the fourth step is an HM represented by a cepstrum region.
The above S for the 0th order term of the average value of the distribution of the output probabilities of M
This is addition of components including / N. 7. The method of creating a hidden Markov model according to claim 5, wherein the fourth step is H represented in a logarithmic spectral domain.
It is the addition of the component including the above S / N to the average value of the distribution of the output probability of MM. 8. The hidden Markov model creation method according to claim 1, wherein said first step comprises a sixth step of converting each output probability distribution of said speech HMM and said noise HMM into a linear spectral domain, and said linear spectral domain The seventh step of multiplying the output probability distribution of the speech HMM by the above-mentioned multiplicative distortion, the distribution of the output probability of the speech HMM multiplied by the above-mentioned multiplicative distortion, and the distribution of the output probability of the noise HMM in the above-mentioned linear spectral region are described. It comprises an eighth step of performing a convolution operation and a ninth step of inversely transforming the result of the convolution operation into the region of the original speech HMM. 9. The method of creating a hidden Markov model according to claim 1, wherein the first step is to convert a distribution of output probabilities of the speech HMM and the noise HMM into a linear spectral domain.
And S / N in the distribution of the output probability of one of the speech HMM and the noise HMM in the linear spectral domain.
And a seventh step of convoluting the output probability distribution of the multiplied HMM and the output probability distribution of the non-multiplied HMM, and a result of the convolution calculation of the original speech HMM. (9) Step 9 of inverse transformation into the region of. 10. The method of creating a hidden Markov model according to claim 8 or 9, wherein in the sixth step, the distributions of the output probabilities of the speech HMM and the noise HMM respectively represented in the cepstrum domain are cosine-transformed and further extracted. The ninth step is a step of performing a logarithmic transformation of the convolution operation result, and further an inverse cosine transformation of the result. 11. The hidden Markov model creation method according to claim 8 or 9, wherein in the sixth step, the distributions of the output probabilities of the speech HMM and the noise HMM respectively expressed in the logarithmic spectral domain are subjected to the exponential transformation. Is the step
The ninth step is a step of logarithmically converting the convolution operation result. 12. The hidden Markov model creation method according to claim 1, wherein said first step comprises a fourth step of multiplying said multiplicative distortion by an output probability distribution of a speech HMM expressed in a linear spectral domain, and said multiplication. The fifth step comprises a convolution operation of the output probability distribution multiplied by the sexual distortion and the distribution of the output probability of the noise HMM expressed in the linear spectral domain. 13. The method of creating a hidden Markov model according to claim 1, wherein the first step comprises a distribution of output probabilities of one of the speech HMM and the noise HMM expressed in a linear spectral domain, the component including the S / N. And a fifth step of performing a convolution operation of the above-mentioned multiplied output probability distribution and the non-multiplied output probability distribution. 14. The hidden Markov model creating method according to claim 5, 9, or 13, wherein the component containing S / N is 10 ^{((S / N) / 2)} . 15. The method of creating a hidden Markov model according to claim 1, wherein the second step is estimated by iterative calculation by a maximum likelihood estimation method. 16. The hidden Markov model creation method according to claim 1, wherein the second step is estimated by iterative calculation by the steepest descent method. 17. The method of creating a hidden Markov model according to claim 1, wherein the second step applies multiplicative distortions of various values to each of the incomplete noise-resistant speech HMMs. Uncompleted noise resistant voice H
The likelihoods of the MM with respect to the input speech are calculated, and the value of the multiplicative distortion given to the HMM having the maximum likelihood is used as the estimated value. 18. The hidden Markov model generation method according to claim 1, wherein the second step gives S / Ns of various values to each of the incomplete noise-resistant speech HMMs. , The likelihood of each of these uncompleted noise resistant speech HMMs with respect to the input speech is obtained, and the S / N value given to the HMM having the maximum likelihood is used as an estimated value. 19. The method of creating a hidden Markov model according to claim 1, wherein the noise HMM used in the first step is created by recording noise under the environment of the speech to be recognized. Have steps. 20. The hidden Markov model generation method according to claim 1, wherein the speech HMM is expressed in a cepstrum domain, and the noise HMM is expressed in a logarithmic spectrum domain. It is a thing. 21. A noise resistant speech H created from a speech HMM not affected by noise and multiplicative distortion and a noise HMM.
A voice recognition device using MM, comprising: a voice HMM storage unit storing the set of voice HMMs; a noise HMM storage unit storing the noise HMMs; and an endurance store storing the set of noise resistant voice HMMs. A noise voice HMM storage unit, a multiplicative distortion or S / N storage unit storing an unknown number of multiplicative distortions or S / Ns, a noise input means for recording noises in the same environment as the recognition target voice, and the above input means. Noise HMM creating means for creating the noise HMM based on the recorded noise and storing it in the noise storage unit, the noise HMM in the noise HMM storage unit, the voice HMM in the voice HMM storage unit, and the multiplyability. Multiplying strain or S of the above unknown in the strain or S / N storage
/ N to create an incomplete noise resistant speech HMM containing multiplicative distortion or S / N as an unknown number
M creating means, voice input means for inputting a voice to be recognized, and estimating the above-mentioned multiplicative distortion or S / N that maximizes the likelihood of the incomplete noise resistant voice HMM for the voice input from the voice input means. Multiplying distortion or S / N estimating means and multiplying distortion or S / N estimated by the above estimating means
To the uncompleted noise-resistant speech HMM to create a set of the noise-resistant speech HMM and store it in the noise-resistant speech HMM storage unit, and a noise-recognition speech HMM completion means for recognizing the speech input by the speech input means. And a voice recognition means for calculating the similarity between the power voice and each of the noise resistant voice HMMs in the noise resistant voice HMM storage unit and outputting the recognition result based on the calculation result.