JPH11212587A - Noise adapting method for speech recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decrease the process quantity of noise adaptation by limiting a distribution of addition of noise HMM in a probability distribution constituting phoneme HMM generated in small-noise environment. SOLUTION: A distribution selection part 302 selects a distribution for model composition with noise HMM out of (n) representative distributions stored in a representative distribution storage part 104 and outputs it to a spectrum conversion part 303. The representative distributions are selected on the basis of distances calculated between one representative distribution and a noise superposition representative distribution obtained by adding the noise HMM to the representative distribution in a frequency range, distribution by distribution. On the basis of the inter-distribution distances of respective previously calculated representative vectors, the distribution selection part 302 stores code words of distributions having inter-distribution distance values exceeding a certain threshold value and the noise HMM is added only to the representative distributions corresponding to the code word in the frequency range. The process quantity of a noise adapting method by the superposition of the noise HMM is nearly proportional to a distribution to which the process is applied. Here, the distribution selection part 302 limits the number of representative distributions which are processed.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a noise adaptation method in speech recognition.

[0002]

2. Description of the Related Art In order to put a speech recognition apparatus into practical use, it is essential to have a noise-resistant technique for correctly recognizing speech uttered under noise. In the field of speech recognition, a feature parameter sequence of speech divided into phoneme units created in a noise-free environment is represented by a hidden Markov model (phoneme HMM), and a feature parameter sequence of noise is represented by a hidden Markov model ( A noise adaptive method of adding a noise HMM on a spectrum (Japanese Patent Laid-Open No. 6-214592) is known.

[0003]

In the above-described conventional noise adaptation method, it is necessary to convert all acoustic parameters of the phonological HMM into parameters in the spectral domain, so that the adaptation process takes time. If the noise adaptation processing takes too much time, it is necessary to wait until the noise adaptation is completed when using the speech recognition function, and the usability deteriorates.
Further, in an environment where noise constantly changes, a difference occurs between the noise adaptation time and the time point when speech recognition is actually started, so that the effect of noise adaptation cannot be sufficiently exhibited.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a noise adaptation method in speech recognition that significantly reduces the processing amount in noise adaptation processing by superposition of a noise HMM, and a noise-resistant speech recognition apparatus that operates stably even under noise. Is to provide.

[0005]

In order to achieve the above object, according to the present invention, a feature parameter sequence of a speech divided into phoneme units is represented as a hidden Markov model (phoneme HMM). , And adding the phoneme HMM created in the noise-free environment and a feature parameter sequence of noise as a hidden Markov model (noise HMM) in the frequency domain of each of the feature parameter sequences. In the noise adaptation method in speech recognition for creating the phoneme HMM adapted to the above, the distribution to which the noise HMM is added is restricted from the probability distributions constituting the phoneme HMM created in the noise-free environment.

(Embodiment) An embodiment of the present invention will be described below.
FIG. 1 is a semi-continuous HMM for explaining an embodiment of the present invention.
1 is a block diagram of a word speech recognition device according to the present invention. In FIG. 1, 101 is a speech input unit, 102 is a feature parameter extraction unit, 103 is a speech recognition unit, 104 is a representative distribution storage unit,
105 is an acoustic HMM storage unit, 106 is a dictionary storage unit, 10
7, a recognition result output unit; and 108, a noise model adaptation unit.

The voice input unit (101) is usually composed of a microphone and an A / D converter, and takes in a voice waveform and converts it into digital data. The feature parameter extraction unit (102) divides the input speech waveform converted into digital data at short time intervals (usually 10 ms to 20 ms), and extracts feature parameters for the divided data. As feature parameters in speech recognition, LPC
Cepstrum is often used. For more information on LPC cepstrum, see Nakata's "Voice" (Corona). The phonetic HMM is stored in the acoustic HMM storage unit (105). The phoneme HMM represents a speech feature parameter sequence divided into phoneme units as a hidden Markov model (HMM). For HMM, Nakagawa, "Speech Recognition by Probabilistic Model" (IEEJ)
A detailed description is given in The probability density function in the phonological HMM is approximated as a mixture distribution of a multidimensional Gaussian distribution. At this time, in the semi-continuous HMM method, the multidimensional Gaussian distribution is classified into a finite number of clusters, and a distribution belonging to each cluster is represented by one distribution. Here, the represented distribution is called a representative distribution, a cluster identifier (ID) is called a codeword, and a correspondence table that associates a codeword with a representative distribution of the cluster is called a codebook. This codebook is stored in the representative distribution storage unit (104). One embodiment of the representative distribution storage unit (104) is shown in FIG. When the representative distribution is modeled as a multidimensional Gaussian distribution, it can be represented by a mean vector and a variance matrix. Representative distribution storage unit (104)
Is a mean vector V representing the representative distribution and a variance matrix Σ,
And an ID number (codeword) for specifying the distribution. In the dictionary storage unit (105), the vocabulary to be recognized is described as a phoneme symbol string. The word HMM is created by connecting the phoneme HMM to each recognition target word in accordance with the phoneme symbol string described in the dictionary storage unit (105). The speech recognition unit (103) includes a speech H created based on the time-series data of the feature parameters of the input speech extracted by the feature parameter extraction unit (102) and the vocabulary to be recognized.
The likelihood is calculated with respect to the MM, and the word having the highest likelihood is set as a recognition result. The recognition result output unit (106) is configured by a device that outputs the speech recognition result obtained by the speech recognition unit (103). For example, a display for displaying a recognition result as text corresponds to this.

FIG. 3 is a diagram for explaining an embodiment of the noise model adapting section (108). In FIG.
Denotes a noise HMM creation unit, 302 denotes a distribution selection unit, 303 denotes a spectrum conversion unit, 304 denotes a parameter addition unit, and 305 denotes a spectrum inverse conversion unit.

A noise HMM creating section (301) receives a feature parameter sequence corresponding to a noise waveform to be referred to and creates a noise HMM. The calculation of the input feature parameter sequence has already been described with reference to FIG. Noise HMM
Is a model with a single Gaussian distribution with one appearance probability.
The noise HMM creating unit (301) calculates the average value and the variance of the feature parameters for the input sequence. The distribution selection unit (302) selects a distribution that performs model synthesis with the noise HMM from the representative distributions stored in the codebook storage unit. Details of the distribution selection unit (302) will be described later. The spectrum converter (303) converts the characteristic parameters of the representative distribution selected by the distribution selector (302) into parameters in the frequency domain. As described above, in speech recognition, LPC cepstrum is often used as a feature parameter. The frequency conversion of the LPC cepstrum is performed in two steps: conversion to a logarithmic spectral domain by cosine conversion, and conversion to a linear spectral domain by linear conversion. Here, in the Gaussian distribution of the feature parameter, if the value of the r-th dimension of the average vector is μ [r] and the value of the r-th row and s-column of the variance matrix is σ [r; s], the cosine transform is Given by

[0010]

(Equation 1)

[0011]

(Equation 2)

Here,

[0013]

(Equation 3)

Is a cosine transform matrix. Μ, σ
The subscripts cp and lg of the subscript suffix attached to indicate that they are parameters in the cepstrum domain and parameters in the logarithmic spectrum domain, respectively. Similarly, the linear transformation is performed by the following equation.

[0015]

(Equation 4)

[0016]

(Equation 5)

However, suffix ln indicates that it is a parameter in the linear spectrum region. The parameter adding unit (304) adds each characteristic amount of the representative distribution and the noise HMM on the spectral domain.

[0018]

(Equation 6)

[0019]

(Equation 7)

Here, s of superscript suffix added to μ and σ
n, s, and n indicate the added feature, the feature of the representative distribution, and the feature of the noise HMM, respectively. The spectrum inverse transform unit (305) inversely transforms the characteristic amount of the representative distribution added to the noise HMM from the frequency domain to the original cepstrum domain. This transformation is divided into logarithmic transformation and cosine inverse transformation, which correspond to linear transformation inverse transformation and cosine transformation inverse transformation, respectively. The feature parameters returned to the original cepstrum area again by the spectrum inverse transform unit (305) are stored in the representative distribution storage unit (104), and are used in the speech recognition unit (103) instead of the conventional representative distribution.

Next, details of the distribution selecting section (302) will be described. The distribution selection unit (302) selects a distribution for performing model synthesis with the noise HMM from the n representative distributions stored in the representative distribution storage unit (104), and outputs the selected distribution to the spectrum conversion unit (303). As described with reference to FIG. 2, the representative distributions existing in the representative distribution storage unit (104) are assigned ID numbers (codewords) indicating the respective distributions. Therefore, the distribution selection unit (302) may store the ID number of the representative distribution to be selected.

Hereinafter, a method of selecting a representative distribution performed by the distribution selecting section (302) will be described. The selection of the representative distribution is performed by selecting one of the representative distributions and adding the noise HMM to the representative distribution.
Is calculated in the frequency domain, and the distance between the distribution and the noise superimposed representative distribution is calculated for each distribution, and the calculation is performed based on the magnitude. FIG. 4 illustrates the calculation of the distance between distributions. This figure shows an example of the representative distribution of the code word i among the n representative distributions. This representative distribution i (40
By adding the noise HMM (402) to 1) in the spectral domain, a noise superimposed representative distribution i (403) is created. The noise HMM to be used is created based on environmental noise which is assumed to be used frequently by the speech recognition system.
At this time, the amount of deformation that the representative distribution of the codeword i receives due to the noise superposition is obtained as the distance between the representative distribution (401) and the noise superimposed representative distribution (403). The distance between distributions is Kullbaq divergence (Kullba
The ckdivergence distance measure is often used. The Kullback divergence distance for an uncorrelated multidimensional Gaussian distribution is calculated by the following equation.

[0023]

(Equation 8)

[0024]

(Equation 9)

As described above, the inter-distribution distance Di corresponds to the amount of deformation of this representative distribution due to noise superposition. That is, a representative distribution having a large value is considered to be a model susceptible to noise, and a representative distribution having a small value is considered to be a model robust to noise. Therefore, for a distribution that is robust to noise, noise adaptation processing by superimposing a noise HMM is not performed. Distribution selection unit (30
In 2), based on the value of the inter-distribution distance of each representative vector calculated in advance, a codeword of a distribution in which this value exceeds a certain threshold is stored. Then, the noise HMM is added in the frequency domain only to the representative distribution corresponding to the stored codeword. The processing amount of the noise adaptation method by superimposing the noise HMM is almost proportional to the number of distributions to which the processing is applied. Therefore, by limiting the number of representative distributions to be processed in the distribution selection unit (302), it is possible to greatly reduce the processing amount of noise adaptation. For example, if the number of distributions to be processed is limited to １ ／, the noise adaptation processing is also reduced to approximately 処理. In addition, by employing a distribution distance scale corresponding to the amount of deformation of a model received by noise superposition as a selection criterion, the effect of noise adaptation does not significantly deteriorate due to the reduction in processing amount.

In the first embodiment, the distribution selector (302)
Has described the method of limiting the number of representative distributions for adding the noise HMM in the frequency domain. However, there is a case where the processing time for noise adaptation has a margin depending on the use situation. Therefore, as a second embodiment, when it is necessary to process noise adaptation at high speed, the number of representative distributions to be processed is limited in the same manner as in the first embodiment. A method of superimposing the noise HMM on the representative distribution will be described. In the second embodiment, the distribution selection unit (30
In 2), a priority table as shown in FIG. 5 is prepared. In the priority table, a representative distribution storage unit (104)
And their priorities. Here, the priority is determined based on the value of the distance between distributions. The noise model adaptation unit (108) selects a representative distribution in descending order of priority, and sequentially processes the noise HMM and k superposition. At this time, even if speech recognition is started before the process of noise adaptation is completed, since the representative distribution having a high priority (which is easily deformed by noise) has already been superimposed on the noise HMM, The performance of noise adaptation is not significantly reduced.

As a similar idea, a method of changing the frequency of updating noise adaptation for each representative distribution according to the priority may be considered. In other words, the higher the priority distribution, the more frequently the update is performed, and the lower the priority distribution, the lower the update frequency, so that effective noise adaptation processing can be performed within a limited processing time. Become.

Although all of the embodiments have been described with reference to a word speech recognition apparatus, the present invention can be similarly applied to continuous speech recognition.

[0029]

As described above, according to the present invention, of the probability distributions constituting the phoneme HMM created in the noise-free environment, the distribution to which the noise HMM is added is limited. The processing amount can be significantly reduced. That is, the processing amount of the noise adaptation method by superimposition of the noise HMM is substantially proportional to the number of distributions to which the processing is applied. Therefore, if the number of distributions to be processed is reduced, the processing amount of the noise adaptation can be reduced.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a speech recognition device.

FIG. 2 is a diagram illustrating a representative distribution storage unit.

FIG. 3 is a configuration diagram of a noise model adaptation unit.

FIG. 4 is a diagram for explaining calculation of a distance between distributions.

FIG. 5 is a diagram illustrating a priority table.

[Explanation of symbols]

102: feature parameter extraction unit; 104: representative distribution storage unit; 301: noise HMM creation unit; 302: distribution selection unit;
303: spectrum conversion unit; 304: parameter addition unit.

Claims (4)

[Claims] 1. A method in which a feature parameter sequence of a speech divided into phoneme units is expressed as a hidden Markov model (phoneme HMM) in a noise-free environment, and the phoneme HMM and the noise created in the noise-free environment are created. Expressed as a hidden Markov model (Noise HM
M) in the frequency domain of each of the feature parameter sequences, and a noise adaptation method in speech recognition for creating the phoneme HMM adapted to a vocal environment, wherein the phoneme HMM created in the noise-free environment is configured. A noise adaptation method in speech recognition, wherein a distribution to which the noise HMM is added is restricted from a probability distribution. 2. The phoneme H created in the noise-free environment.
A probability distribution forming the MM, the probability distribution and the noise HM
M and the noise superimposition probability distribution obtained by adding in the frequency domain, the distribution distance is calculated in advance for each distribution, and the distribution to which the noise HMM is added is limited based on the magnitude of the distribution distance. 2. The noise adaptation method in speech recognition according to claim 1, wherein: 3. The distribution to which the restricted noise HMM is added is a distribution in which the value of the distance between the distributions calculated in advance exceeds a predetermined threshold. Noise adaptation method in speech recognition of humans. 4. The processing according to claim 2, wherein the processing of adding said noise HMM is performed with priority given to the distribution having a large value of said distance between distributions calculated in advance. Noise adaptation method in speech recognition as described.