JP2000075888A - Learning method of hidden markov model and voice recognition system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a learning method of a hidden Markov model(HMM) capable of achieving a higher recognition rate under noise environments by performing HMM learning from voice data for learning, outputting a noise mixed voice model, removing an estimated noise model, and outputting a non- noise-dependent voice model with noise removed. SOLUTION: A noise mixed voice data 10 for learning are voice data mixed with noise. An HMM learning processing part 11 performs HMM learning based on the noise mixed voice data 10 for learning, and outputs the result to a noise subtraction processing part 15 as a noise mixed voice model 12. A noise estimation part 13 estimates noise from the noise mixed voice data 10 for learning and outputs an estimated noise model 14. The noise subtraction processing part 15 removes the estimated noise model 14 from the noise mixed voice model 12, and outputs a non-noise-dependent voice model 16 with noise removed. As a result, a sturdy voice model can be prepared without being confined to a specific noise at the time of learning.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method for a Hidden Markov Model and a speech recognition system using the same. For example, the present invention relates to a noise environment such as a car environment used in car navigation or the like which can be operated by voice. The present invention relates to a suitable Hidden Markov Model learning method and a speech recognition system below.

[0002]

2. Description of the Related Art As a speech recognition technique, the classical pattern matching technique has been changed to a statistical technique in recent years, and the latter is becoming mainstream. In the latter statistical method, a Markov model having a probabilistic finite state has been proposed, and is usually referred to as an HMM (hidden Markov model: hidden Markov model). In the HMM, it is possible to increase a high recognition rate by learning a speech model using learning speech data.

FIG. 4 is a block diagram showing the configuration of a conventional continuous speech recognition system using this type of HMM.

In FIG. 4, a continuous speech recognition system comprises:
A / D conversion unit 1, LPC analysis unit 2, background noise sequential learning unit 3, voice detection unit 4, switching unit 5, Viterbi matching unit 6, HM
It comprises an M parameter estimator 7 and an HMM speech dictionary 8. The A / D converter 1 and the LPC analyzer 2
Constitutes a speech analysis block, the background noise sequential learning section 3 and the speech detection section 4 constitute a speech section detection block, and the switching section 5 and the Viterbi comparison section 6 constitute an HMM collation block.
An HMM model learning block is composed of the M parameter estimating unit 7 and the HMM speech dictionary 8.

The A / D converter 1 samples an input audio signal at a predetermined sampling frequency (for example, 8 kHz) and converts it into a digital signal.

[0006] The LPC analysis unit 2 divides a speech waveform into short sections (referred to as frames, whose length is usually 10 to 30 ms), and extracts characteristic parameters for each frame. For speech analysis, LPC (Linear Predictive Code), which is widely used as an efficient method suited to the characteristics of speech, is used.
ng: Linear Predictive Coding) analysis using LPC coefficients to LP
Calculate Cepstrum. Here, the cepstrum is obtained by performing an inverse Fourier transform of a logarithmic spectrum (Logarithm), has a property close to human auditory characteristics, and can express speech efficiently with a relatively small number of parameters.

The speech detector 4 stores the logarithmic power in the noise section and the estimated average value of the LPC cepstrum as a noise feature spectrum, finds the distance between the noise feature vector and the feature vector of the input signal, and calculates the distance from the time change. Detect voice section.

[0008] The background noise sequential learning unit 3 updates the noise feature vector in the section determined to be a noise section,
In addition to performing the sequential adaptive learning of the noise feature, the threshold is also automatically set by the adaptive learning of the distance variation.

The Viterbi collation unit 6 performs HMM collation using the Viterbi algorithm. In the HMM collation, a similarity is obtained by comparing an HMM model representing a phoneme or a word with an unknown input speech.

[0010] The HMM parameter estimating unit 7 generates an EM (Expe
HMM model learning is performed using an ctation Maximization algorithm. In the HMM model learning, the parameters of the HMM model are estimated from voice data prepared in advance.

The switching unit 5 switches the processing between the HMM collation and the HMM model learning. Also, HM
The M speech dictionary 8 stores the HM
The M model learning result is stored, and the HM
Referenced in M matching.

In general, an HMM has multiple states (eg,
And the transition between states. Furthermore, HM
M has a transition probability representing transition between states and an output probability distribution (usually using a Gaussian distribution) for outputting a feature vector accompanying the transition. FIG. 5 shows an example of word speech recognition using such an HMM.

FIG. 5 shows a word H used in the speech recognition method.
It is a state transition diagram which shows the structure of MM.

In FIG. 5, s1, s2, s3, and s4 represent states of voice features in the HMM, and a11, a12, a22, a
23, a33, a34, a44, a45 are state transition probabilities, (u1,
.sigma.1), (u2, .sigma.2), (u3, .sigma.3), and (u4, .sigma.4) represent output probability distributions.

In the HMM, the state transition probability aij (i = 1,
, 4, j = 1,..., 5)
The vector is output using the output probability distribution (uk, σk). In order to recognize an uttered word using the HMM, the HMM is first learned by using learning data prepared for each word so as to output a vector sequence of the word with the highest probability. Next, the vector sequence of the uttered unknown word is input, and the word HMM that gives the highest output probability is set as the recognition result.

In this type of speech recognition method, an uttered word itself is given an HMM and learned, and the likelihood (ie,
The recognition result is determined based on the output probability of the vector train). Although such a word HMM guarantees excellent recognition accuracy, it has disadvantages such as an enormous amount of learning data required due to an increase in the number of recognition vocabularies, and the inability to recognize speech other than the learning target words at all. is there.

Incidentally, in the case of speech recognition in a special environment where noise is large, for example, when a place name is input by voice recognition in car navigation, the performance is significantly deteriorated in a very noisy environment such as in a car, which is accepted by the user. Achieving recognition rates is difficult. However, H
According to the MM theory, it is conceivable to prepare learning speech data mixed with noise and perform speech model learning by HMM learning from this. Hereinafter, this method is called a noise learning method.

By using the voice model obtained in this way, a considerable recognition rate can be obtained even when used in an actual field. However, this noise learning method has a serious limitation that the noise environment of the training speech data must be the same as that when using the speech recognition system.
That is, since the speech model obtained by learning depends on the noise at the time of learning, the recognition rate deteriorates if the noise environment is different. Therefore, it is not practical to apply this method in a real field.

On the other hand, many noise countermeasures have been proposed in the past, such as a noise removal method called spectrum subtraction and adaptive processing using a plurality of microphones.
Recently, a method of adapting the HMM model to noise, called the ΡMC (Parallel Model Combination) method, has been studied well. For example, IEICE Technical Report SP92-96, Frank Mart
in, Kyohiro Shikano, Yasuhiro Minami, Yoichi Okabe: "R
ecognition of NoisySpeech by Composition of Hiddon
Markov Models ".

According to the ΡMC method, a speech model that does not depend on noise is not used, so that it can be applied to any noise environment.

[0021]

However, such a conventional noise-tolerant speech recognition system has the following problems.

For example, when an attempt is made to recognize speech in a car,
Although the PMC method provides some improvement in the recognition rate, it can still achieve a recognition rate that is still insufficient for use in the actual field. The reason is that the ΡMC method mainly targets additive noise, but it is considered that factors such as multiplicative noise and deformation of the utterance spectrum itself under noise called Lambert effect exist in the actual field.

The present invention expands the ΡMC method so as to cope with multiplicative noise, Lambert effect, and the like that are not handled by the conventional PMC method, and to achieve a higher recognition rate in a noise environment.・ It is an object of the present invention to provide a Markov model learning method and a speech recognition system.

[0024]

A learning method for a Hidden Markov Model according to the present invention performs Hidden Markov Model (HMM) learning from learning speech data mixed with noise to generate a speech model. A method for learning a Markov model, which estimates noise included in training speech data, generates an estimated noise model, performs HMM learning from the training speech data, and outputs a noise mixed speech model; The method is characterized in that an estimated noise model is removed from a noise mixed speech model and a noise-independent speech model from which noise has been removed is output.

The above-described estimation noise model is removed by converting the output probability distributions of the estimation noise model and the noise-mixed speech model from values in the cepstrum space to values in the logarithmic spectrum space by cosine transform, and further to linearity by exponential transformation. The mean and variance of the output probability distribution of the estimated noise model in the linear spectrum space are converted to the values in the spectrum space, and the mean and variance of the output probability distribution of the noise mixed speech model in the linear spectrum space calculated in this way are respectively. Is subtracted, and the result is logarithmically converted to a value in logarithmic spectrum space, and an output probability distribution in cepstrum space is obtained by inverse cosine transform.Thus, additive noise is removed from the noise mixed speech model. It may be performed by obtaining a speech model.

Also, the speech recognition system according to the present invention
Hidden Markov model (HM
M) In a speech recognition system that creates a speech model by performing learning and performs speech recognition using the speech model, a speech database for learning that includes speech data for learning mixed with noise, and a speech database that is included in the learning speech data Noise estimating means for estimating noise to generate an estimated noise model; HMM learning processing means for performing HMM learning from training speech data to output a noise mixed speech model; and removing the estimated noise model from the noise mixed speech model. And a noise subtraction processing means for outputting a noise-independent speech model from which noise has been removed, and a speech recognition means for performing speech recognition using the noise-independent speech model.

In the speech recognition system according to the present invention, the noise subtraction processing means converts each output probability distribution of the estimated noise model and the noise mixed speech model from a value in cepstrum space to a value in logarithmic spectrum space by cosine transform. And
Furthermore, the output probability of the estimated noise model in the linear spectrum space is converted into a value in the linear spectrum space by exponential conversion, and the mean and variance of the output probability distribution of the noise mixed speech model in the linear spectrum space calculated in this way are respectively obtained. The mean and variance of the distribution are subtracted, the result is logarithmically converted to a value in logarithmic spectrum space, and the output probability distribution in cepstrum space is obtained by inverse cosine transform, and thus, among the noise mixed speech models, A speech model from which the objective noise has been removed may be obtained.

A speech recognition system according to the present invention synthesizes a noise-independent speech model, noise estimation means for estimating noise from an input to generate an estimated noise model, and a noise-independent speech model and an estimated noise model. And a model synthesizing unit that outputs a noise-mixed speech model, and a recognition unit that performs an HMM recognition process using the noise-mixed speech model.

The model synthesis means may synthesize the noise-independent speech model and the noise model by the PMC method.

[0030]

Embodiments of the present invention will be described below with reference to the drawings.

First Embodiment FIG. 1 is a block diagram showing an apparatus used for implementing a method for learning a Hidden Markov Model according to a first embodiment of the present invention.

In FIG. 1, reference numeral 10 denotes learning noise mixed voice data (learning voice data) provided as a voice database. The learning noise-mixed speech data 10 is speech data in which noise is mixed.

The HMM learning processing unit 11 (HMM learning processing means) performs HMM learning based on the learning noise-mixed speech data 10 by a conventional method, and outputs the result as a noise-mixed speech model 1
As 2 is output to the noise subtraction processing unit 15 (noise subtraction processing means).

The noise estimating unit 13 (noise estimating means) estimates noise from the noise mixed speech data 10 for learning and outputs an estimated noise model 14.

The noise subtraction processing unit 15 removes the estimated noise model 14 from the noise mixed speech model 12 and outputs a noise-independent speech model 16 from which noise has been removed.

As described above, the apparatus used for implementing the hidden Markov model learning method according to the present embodiment is as follows.
An HMM learning processing unit 11 that performs ΗMM learning based on the learning noise-mixed speech data 10 and outputs the result as a noise-mixed speech model 12, and removes noise included in the learning noise-mixed speech data 10 when performing HMM learning on the speech model. A noise estimating unit 13 for estimating and generating an estimated noise model 14; a noise subtracting unit 15 for removing the estimated noise model 14 from the noise mixed speech model 12 and outputting a noise-independent speech model 16 from which noise has been removed; It is composed of

The operation of the Hidden Markov Model learning method performed by the above-configured apparatus will be described below.

The learning noise mixed speech data 10 is a speech database in which noise is mixed. The HMM learning processing unit 11 performs HMM learning based on the learning noise-mixed speech data 10 by a conventional method. The resulting speech model is referred to as a noise mixed speech model 12. This noise-mixed speech model 12 is used for learning noise-mixed speech data 10.
This model reflects the nature of the noise mixed into the sound, and by using this model, speech with the same nature can be recognized at a high recognition rate, but noise of a different nature can be recognized. There is a feature that it is weak to the voice that is.

On the other hand, the noise estimation section 13 estimates noise from the noise mixed speech data 10 for learning, and
4 is output. Here, a one-state, one-mixed HMM model is considered as the estimated noise model 14. Therefore, only the mean and the variance are used as model parameters.
As a method of creating the estimated noise model, for example, a non-utterance section is extracted from the noise-mixed speech data for learning 10, a signal in the section is regarded as noise, and its average and variance may be obtained.

The above-described noise-mixed speech model 12 is roughly described as follows.
Hereinafter, it is considered that the voice / noise of (A), (B) and (C) is integrated.

(A) Speech deformed by the Lambert effect (Β) Multiplicative noise (C) Additive noise This noise-mixed speech model 12 has a low recognition rate for speech mixed with noise different from that at the time of learning. Is the above (C)
This is probably due to the large difference in the nature of the additive noise. That is, the above-mentioned (B) multiplicative noise is mainly a problem of a sound transmission system such as a microphone, and has no relation to noise inside and outside the vehicle if the microphone is the same. Although the Lambert effect is considered to depend to some extent on the type and degree of noise, it is considered that the effect of the noise difference is relatively small as compared with the additive noise (C).

The purpose of the noise subtraction processing unit 15 is (A) the sound transformed by the Lambert effect obtained by learning, and
(B) The characteristics vary greatly depending on the type and degree of the noise while leaving the multiplicative noise. (C) By removing the additive noise, it is robust against the Lambert effect and the multiplicative noise, and has a specific type or noise. The goal is to generate a speech model that is not limited to a degree of additive noise. The speech model thus generated is referred to as a noise-independent speech model 16 here. Of course, this speech model alone cannot cope with additive noise, but this can be coped with by applying the conventional method of MC processing. Details of this will be described in a second embodiment.

FIG. 2 is a flow chart showing the flow of the operation of the noise subtraction processing section 15. In FIG. 2, S indicates each step of the flow. The basic idea of this processing is PMC
Same as the method.

While the MC method adds and combines the noise model and the voice model to generate a voice model in which noise is mixed, the processing in the noise subtraction processing unit 15 uses the noise model from the voice model in which noise is mixed. Is subtracted to generate a speech model from which the additive noise has been removed (but the learning result of the Lambert effect and the multiplicative noise is preserved).

Hereinafter, the operation of the noise subtraction processing unit 15 will be described with reference to FIG. Subscript cp, l of vector M, N, S
g and ln are cepstrum, logarithm spectrum, and linear spectrum (Linea spectrum), respectively.
r) Indicates a value in space.

The HMM output probability distribution {Mcp, {Mcp} (Mcp is the average vector of the output probability distribution, and {Mcp is the variance matrix of the output probability distribution) using the cepstrum as a parameter is calculated from the cepstrum space by the cosine transform in step S1. Value in space ｛Mlg, ΣM
lg}. Further, the values {Mln, {Ml} in the exponential spectrum space are obtained by the exponential conversion in step S2.
n}. Steps S1 and S2 are the same as those of the conventional PMC method, and the details are omitted.

These processes are performed for each of the noise mixed speech HMM and the noise ΗMM.

Step S3 is a processing different from the conventional PMC method. The output probability distribution {Mln, {Mln} of the noise-mixed speech and the output probability distribution {Nln, {Nln} of the noise in the linear spectrum space are expressed by the following equation (1). ) And (2) are subtracted to obtain an output probability distribution {Sln, {Sln} in the linear spectrum space of the noise-free speech.

Sln (k) = max {Mln (k) −Nln (k), α * Nln (k)} (1) {Sln (k, l) = max ｛ΣMln (k, l) −ΣNln (k , L), α * α * {Nln (k, l)} (2) k, l = 1, 2,..., P (p is a vector Sln, Mln
However, for example, Sln (k) represents the k-th component of the average vector Sln of the output probability distribution, and ΣSln (k,
1) represents the k-th and l-th components of the variance matrix ΣSln of the output probability distribution. α is a positive constant, for example, α = 0.
It may be 1.

Equations (1) and (2) can be considered basically as subtracting the noise spectrum from the spectrum of the noise-mixed speech. In this case, in order to prevent the result of the subtraction from becoming a negative value, a value obtained by multiplying the noise spectrum by a constant α is used as the flooring as the minimum value. By this processing, the additive noise component has been removed from the noise mixed speech.

The speech probability distribution {Sln,} on the linear spectrum space thus obtained from which additive noise has been removed.
Sln} is logarithmically converted in step S4 to a value {Slg, {Slg} in a logarithmic spectrum space,
Further, in step S5, inverse cosine transform is performed to obtain values {Scp, {Scp} in the cepstrum space. Steps S4 and S5 are the same as the conventional PMC method, and a detailed description thereof will be omitted.

In this way, a speech HMM model from which additive noise has been finally removed is obtained. Since this speech model only removes additive noise and preserves the learning results of the Lambert effect and multiplicative noise, it is robust against the Lambert effect, etc., and has a specific type and degree of additive noise. Can be considered as a speech model that is not limited to. The resulting speech model is output as a noise-independent speech model 16.

As described above, in the learning method of the Hidden Markov Model according to the first embodiment, the HMM learning that performs the ΗMM learning based on the noise-mixed speech data for learning 10 and outputs as the noise-mixed speech model 12. Processing unit 11
And a noise estimating unit 13 for estimating noise included in the training noise-mixed speech data 10 and generating an estimated noise model 14 when the speech model is subjected to HMM learning. A noise subtraction processing unit 15 that outputs a noise-independent speech model 16 from which noise has been removed by removing the noise, estimates noise included in the training noise-mixed speech data 10, and generates an estimated noise model 14. The HMM learning is performed from the noise-mixed speech data for learning 10 to output a noise-mixed speech model 12, and the noise subtraction unit 15 removes the estimated noise model 14 from the noise-mixed speech model 12 to remove the noise. Since the dependent speech model 16 is output, the noise-mixed speech data for learning is obtained from the noise-mixed speech model 12 obtained by the HMM learning. Only additive noise estimated from 0 is to be removed, without being limited to the particular noise during training, it is possible to create a robust speech model for such Lambert effect.

Therefore, if applied to a speech recognition system using an HMM, the $ MC method is extended to
It is possible to cope with multiplicative noise, Lambert effect, and the like that are not handled by the C method, and achieve a higher recognition rate in a noise environment.

Second Embodiment The second embodiment is a system for performing speech recognition using the noise-independent speech model obtained in the first embodiment.

In the first embodiment, only the additive noise estimated from the learning speech data is removed, so that the present invention is not limited to a specific noise at the time of learning and is robust against the Lambert effect and the like. I was able to create a new voice model
However, since the additive noise has been removed from the noise-independent speech model 16, it cannot be used as it is for recognition of speech mixed with additive noise. As described in the description of the related art, various methods have been conventionally proposed to cope with the additive noise.

Therefore, it is conceivable to reduce the additive noise in the input speech by using, for example, a spectral subtraction method, and then perform speech recognition by using the noise-independent speech model of the first embodiment. In the method using the spectral subtraction method, considering that the noise-independent speech model is robust against the Lambert effect, a certain improvement in performance can be expected by using the noise-independent speech model.

However, considering that the noise-independent speech model of the first embodiment is basically generated according to the concept based on the PMC method, the above-described spectral subtraction method is applied as a method for dealing with additive noise. Than
It is clear that the application of the MC method has a better affinity with the noise-independent speech model, and thus the use of the PMC method is most effective in increasing the high recognition rate.

FIG. 3 is a block diagram showing the configuration of the speech recognition system according to the second embodiment of the present invention. In the description of the speech recognition system according to the present embodiment, the same parts as those in the block diagram shown in FIG.

In FIG. 3, reference numeral 20 denotes a voice input unit for converting a voice input from a microphone or the like into a digital signal and inputting the digital signal. Reference numeral 21 divides a voice waveform into short sections, extracts feature parameters for each frame, and analyzes the voice. This is a voice analysis unit including an LPC analysis unit.

The noise estimating unit 13 estimates the noise in the input signal based on the output of the speech analyzing unit 21, and
As 4, it is output to a model synthesis unit (PMC processing unit) 22 (model synthesis means).

The noise-independent speech model 16 is a speech model generated by the noise subtraction processing section 15 (FIGS. 1 and 2). It is a robust and robust speech model that is not limited to any particular type or degree of additive noise.

The model synthesizing section 22 calculates the estimated noise model 14
And a noise-independent speech model 16 by using the PMC method, and outputs a new speech model as a noise-mixed speech model 23. The algorithm of the PMC method is the same as the conventional method.

The analysis result from the speech analysis unit 21 and the noise-mixed speech model 23 are output to a recognition unit 24 (recognition means). The recognition unit 24 performs a recognition process on the output of the model synthesis unit 22 using the noise-mixed speech model 23, outputs a recognition result 25, and ends the process.

As described above, the learning method of the hidden Markov model according to the present embodiment is the same as the conventional ΡMC method except that the noise-independent speech model 16 is used instead of the clean speech model as the speech model. is there.

The operation of the speech recognition system configured as described above will be described below.

The voice input from the voice input unit 20 is input to the voice analysis unit 21, analyzed by the voice analysis unit 21, and output to the noise estimation unit 13 and the recognition unit 24.

The noise estimating unit 13 estimates noise from the input, creates an estimated noise model 14, and outputs it to the model synthesizing unit 22. On the other hand, by removing the additive noise, the model synthesis unit 22 is a noise-independent sound model that is robust to the Lambert effect and multiplicative noise, and is not limited to a specific type or degree of additive noise. An audio model 16 is input.

The model synthesis unit 22 performs PMC synthesis on the estimated noise model 14 and the noise-independent speech model 16,
Noise-mixed speech model 23 considering the effects of additive noise
Get.

The recognizing unit 24 performs a recognizing process using the noise-mixed speech model 23, and outputs a recognition result 25.

ΡM using a conventional clean speech model
Compared with the C method, the PMC method is effective against additive noise, but is ineffective against deformation of the voice itself such as the Lambert effect. On the other hand, in the present method, since the noise-independent speech model is robust against the Lambert effect, it is possible to construct a speech recognition system that is robust against the Lambert effect as well as the additive noise.

Further, as compared with the case where the noise-mixed speech model learned from the noise-mixed speech data for learning 10 is used as it is (noise learning method), the noise-mixed speech model 23 is more robust to the Lambert effect, but has the additive noise. The type and degree of are limited only to the noise at the time of learning, and it is not possible to estimate the additive noise from the input at the time of the test and generate a new noise mixed speech model adapted to the noise as in the モ デ ル MC method. In the present embodiment, it is possible to cope with additive noise different from that at the time of learning by applying the ΡMC method.

As described above, the speech recognition system according to the second embodiment estimates the noise from the input to generate the estimated noise model 14, the noise-independent speech model 16, and the estimated noise model. A model synthesis unit 22 that synthesizes the model 14 and outputs a noise-mixed speech model 23; and a recognition unit 24 that performs HMM recognition processing using the noise-mixed speech model 23. By using the noise-independent speech model 16 in which only additive noise has been removed from the noise-mixed speech model that has been subjected to noise learning, the speech mixed with noise is recognized by the ΡMC method, so that the speech is limited to specific noise. In addition, a robust noise-recognition speech recognition system that is robust not only to additive noise but also to the Lambert effect and the like can be configured. In addition, since the noise-independent speech model 16 has affinity with the PMC method, when recognizing the noise-mixed speech using the noise-independent speech model, by using the ΡMC method together,
Higher recognition performance is obtained than when used in combination with other additive denoising methods.

Therefore, if the speech recognition system having such excellent features is applied to, for example, a noise-resistant speech recognition system used for car navigation which can be operated by speech, the recognition rate can be greatly improved in this device. Can be planned.

The method for learning the Hidden Markov Model according to the first embodiment and the speech recognition system according to the second embodiment may be any speech recognition system that performs HMM learning from training speech data. It can be applied to various types of circuits, and can also be implemented as circuits incorporated in various terminals.

Further, the method of learning the Hidden Markov Model according to the first embodiment and the number of processing units and various processes constituting the speech recognition system according to the second embodiment,
The type connection state and the like are not limited to the above embodiments.

[0077]

In the learning method of the Hidden Markov Model according to the present invention, noise included in the learning speech data is estimated, an estimated noise model is generated, and HMM learning is performed from the learning speech data. Since the noise-mixed speech model is output and the estimated noise model is removed from the noise-mixed speech model by the noise subtraction processing means to output a noise-independent speech model from which noise has been removed, the noise-independent speech model is estimated from the training speech data. Thus, only additive noise is removed, and a speech model that is robust against the Lambert effect and the like can be created without being limited to specific noise during learning.

In the speech recognition system according to the present invention, a learning speech database having learning speech data mixed with noise, and a noise estimating means for estimating noise contained in the learning speech data and generating an estimated noise model. HMM learning processing means for performing HMM learning from training speech data to output a noise-mixed speech model, and noise for removing an estimated noise model from the noise-mixed speech model and outputting a noise-independent speech model from which noise has been removed Since it is configured to include subtraction processing means and speech recognition means for performing speech recognition using a noise-independent speech model, the MC method is extended to
It is possible to cope with multiplicative noise, Lambert effect, and the like that are not handled by the C method, and achieve a higher recognition rate in a noise environment.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an apparatus used for implementing a hidden Markov model learning method according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a flow of an operation of a noise subtraction processing unit in the hidden Markov model learning method.

FIG. 3 is a block diagram illustrating a configuration of a speech recognition system according to a second embodiment to which the present invention has been applied.

FIG. 4 is a block diagram showing a configuration of a conventional continuous speech recognition system using an HMM.

FIG. 5 is a diagram showing a structure of a word Hidden Markov Model used in a speech recognition method.

[Explanation of symbols]

10 noise mixed speech data for learning (speech data for learning), 11 HMM learning processing section (HMM learning processing means), 12, 23 noise mixed speech model, 13 noise estimation section (noise estimation means), 14 estimated noise model, 15
Noise subtraction processing section (noise subtraction processing means), 16 noise-independent speech model, 20 speech input section, 21 speech analysis section,
22 model synthesis unit (PMC processing unit) (model synthesis means), 24 recognition unit (recognition means), 25 recognition result

Claims (6)

[Claims] 1. A Hidden Markov Model (HMM) learning method for performing a Hidden Markov Model (HMM) learning from learning noise data mixed with noise to generate a speech model, wherein the learning voice data is included in the learning voice data. And generating an estimated noise model, performing HMM learning from the learning speech data to output a noise-mixed speech model, and removing the estimated noise model from the noise-mixed speech model to reduce noise. A method for learning a Hidden Markov Model, comprising outputting a noise-independent speech model from which noise has been removed. 2. The method of removing an estimated noise model comprises converting each output probability distribution of an estimated noise model and a noise mixed speech model from a value in a cepstrum space to a value in a logarithmic spectrum space by a cosine transform, and further performing an exponential transform. Is converted to a value in the linear spectrum space, and the average and the variance of the output probability distribution of the noise mixed speech model in the linear spectrum space calculated in this way are respectively averaged for the output probability distribution of the estimated noise model in the linear spectrum space. And the variance, logarithmically transform the result to a value in logarithmic spectrum space, and obtain an output probability distribution in cepstrum space by inverse cosine transform. 2. The method according to claim 1, wherein said step is performed by obtaining a removed speech model. Learning method of Markov model. 3. A Hidden Markov Model (HMM) learning is performed from the learning voice data to generate a voice model,
A speech recognition system for performing speech recognition using the speech model, comprising: a learning speech database having learning speech data mixed with noise; and estimating noise included in the learning speech data to generate an estimated noise model. Noise estimation means; HMM learning processing means for performing HMM learning from the learning speech data to output a noise-mixed speech model; noise removal means for removing the estimated noise model from the noise-mixed speech model to remove noise. A speech recognition system comprising: noise subtraction processing means for outputting a dependent speech model; and speech recognition means for performing speech recognition using the noise-independent speech model. 4. The noise subtraction processing means converts each output probability distribution of the estimated noise model and the noise mixed speech model from a value in cepstrum space to a value in logarithmic spectrum space by cosine transform, and further performs exponential transformation. The average and the variance of the output probability distribution of the estimated noise model in the linear spectrum space are converted to the values in the linear spectrum space, and the average and the variance of the output probability distribution of the noise-mixed speech model in the linear spectrum space calculated in this manner are respectively calculated. Subtract the variance, logarithmically convert the result to a value in logarithmic spectrum space, and obtain the output probability distribution in cepstrum space by inverse cosine transform, thus removing additive noise from the noise mixed speech model 4. The speech recognition system according to claim 3, wherein a speech model obtained is obtained. 5. A noise-independent speech model, noise estimation means for estimating noise from an input to generate an estimated noise model, and synthesizing the noise-independent speech model and the estimated noise model, The speech recognition system according to claim 3, further comprising: a model synthesis unit that outputs a model; and a recognition unit that performs an HMM recognition process using the noise-mixed speech model. 6. The model synthesizing means includes a PMC (Parallel).
The noise-independent speech model and the noise model are synthesized by an el Model Combination (el Model Combination) method.
A speech recognition system as described.