JP2003140684A - Method, device, and program for voice recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and device for voice recognition for recognizing the voice of a speaker other than a leaner of HMM with high precision. SOLUTION: In the voice recognition using an HMM having a mixed multidimensional regular distribution, equal output probability is used for input feature vectors within a specific distance range having its center at the mean vector of the mixed multidimensional regular distribution to prevent the recognition precision of an acoustically different speaker from becoming worse deviating from a range of speech prepared as learning data.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to speech recognition using a hidden Markov model.

[0002]

2. Description of the Related Art In recent years, a method using a Hidden Markov Model (HMM) having a mixed multidimensional normal distribution has been studied and applied as an effective method for speech recognition. Regarding this method, for example,
References (Seiichi Nakagawa, "Speech Recognition by Stochastic Model",
For details, see IEICE, 1998). HM with mixed multidimensional normal distribution described in such references
A voice recognition method using M will be briefly described.

First, a speech is composed of its constituent units (phonemes, syllables,
It is represented by a model consisting of states and arcs connecting them, as shown in FIG. Each state has a state transition probability a indicating the probability of transition from that state to another state.
have ij. Further, each arc has an output probability bij (y) of an acoustic feature that is output at the time of transition. Note that i,
Each j represents a state number, and y represents an acoustic feature.

An HMM having a mixed multidimensional normal distribution gives an output probability bij (y) by the following equation (1).

[0005]

[Equation 1]

Here, λijm represents the branch probability for the m-th multidimensional normal distribution of the transition path from the state i to the state j. Further, μijm and Σijm represent the mean vector and the covariance matrix of the m-th multidimensional normal distribution, respectively.

When recognizing an unspecified speaker with high accuracy using this method, it is necessary to learn these parameters from the data collected for a plurality of speakers in advance.

However, in practice, it is impossible to create learning data that assumes all speakers using the system. Therefore, when used in a real environment, the recognition accuracy may drop significantly depending on the speaker. There is.

The decrease in recognition accuracy will be described with reference to FIG.

In FIG. 3, the normal distribution shown by the solid line is the output probability distribution of a certain phoneme learned from the utterances of a plurality of speakers. On the other hand, the output probability distribution of a speaker having a utterance acoustically different from the speaker collected as the learning data (referred to as speaker A) is represented by a broken line.
In FIG. 3, the number of dimensions of the normal distribution is 1 for simplicity.
The feature vector of the voice uttered by the speaker A is y shown in FIG.
, The output probability calculated from the learned normal distribution is P. This P has a small value with respect to the output probability PT calculated from the actual output probability distribution of the speaker A. That is, the output probability for this phoneme becomes unreasonably small. This means that the recognition accuracy decreases. Therefore, when the training data is insufficient, or when a large number of unspecified speakers use the system, the learned output probability distribution covers the acoustic distribution of all possible speakers. Therefore, a speaker whose recognition accuracy is significantly lowered may appear.

To solve this problem, a speaker adaptation technique is used,
A method of adaptively learning a mean vector and a covariance matrix for each speaker has been proposed. However, in order to re-learn these parameters by speaker adaptation, a certain amount of data is required for each speaker, which is not suitable for situations where the speaker changes one after another in a short time.

As described above, in the conventional speech recognition using the HMM having the mixed multidimensional normal distribution, when the prepared data for learning is insufficient, the utterance deviates from the learning data. There is a problem that the recognition accuracy for the speaker is not obtained and it is not possible to deal with an unspecified speaker.

[0013]

In speech recognition using the HMM, it is practical that the required recognition accuracy can be obtained even for a speaker having a utterance acoustically different from the utterance prepared as learning data. It is also effective in terms of aspect.

The present invention has been made in consideration of the above circumstances, and provides a voice recognition method, device, and program capable of performing highly accurate voice recognition even for a speaker who is not an HMM learning target. The purpose is to do.

[0015]

A speech recognition method according to the present invention detects a vocal section of a voice from an input speech signal and analyzes a voice signal of each detected vocal section to extract a feature vector sequence. Then, the extracted feature vector series, and prepared in advance for each predetermined recognition candidate, to calculate the matching score by performing pattern matching with a hidden Markov model having a normal distribution, based on the calculated matching score In the speech recognition method for determining the recognition result of the speech, the pattern matching includes a likelihood calculation of the feature vector series based on the hidden Markov model, and the likelihood calculation is performed from an average vector of the hidden Markov model. It is characterized in that equal output probabilities are given to feature vector sequences within a predetermined distance range.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. Embodiments of the present invention relate to speech recognition using a Hidden Markov Model (HMM) having a mixed multidimensional normal distribution, and a matching likelihood that is robust against a variation in the multidimensional normal distribution depending on a speaker. The present invention relates to a voice recognition device for calculating.

[First Embodiment] FIG. 1 shows a first embodiment of the present invention.
3 is a block diagram showing a schematic configuration of a voice recognition device according to the exemplary embodiment. FIG. The speech recognition apparatus shown in FIG. 1 includes a vocal section detecting unit 1 for acoustically analyzing an input speech signal to detect a vocal section.
01, a feature vector extraction unit 102 that extracts a feature vector by analyzing the speech signal of the utterance section detected by the utterance section detection unit 101, and standard features of each predetermined recognition candidate that has been learned in advance. Hidden Markov Model (HMM) with mixed multidimensional normal distribution as pattern
The standard feature pattern storage unit 104 storing the feature vector sequence, the feature vector series extracted by the feature vector extraction unit 102, and the HMM for each recognition candidate stored in the standard feature pattern storage unit 104 are subjected to pattern matching to perform matching. A pattern matching unit 103 that calculates a score (likelihood calculation) and a recognition result determination unit 105 that determines the recognized utterance content based on the matching score for each recognition candidate obtained by the pattern matching unit 103. Consists of In the configuration shown in FIG.
A microphone for inputting the voice uttered by a speaker and converting it into a digital electric signal (digital voice signal),
A voice input unit including an A / D (analog / digital) converter may be connected. In this case, the vocalization section detection unit 101
The input audio signal to the audio input device is supplied from the audio input unit.

A voice recognition process by the voice recognition apparatus according to the present embodiment having the above-described configuration will be described.

First, the utterance section detection unit 101 detects a utterance section of a voice from an input voice signal. The feature vector extraction unit 102 frequency-analyzes the voice signal of the utterance section detected by the utterance section detection unit 101 for each of a plurality of predetermined frequency bands. As a result, the voice signal in the utterance section is feature vector series (feature vector time series)
Converted to {xt}. The feature vector (feature parameter) is obtained in units of fixed time length called a frame.

As typical feature vectors used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum count obtained by LPC (linear prediction) analysis, etc. are well known. Also in this embodiment, a known feature vector can be used. That is, the present invention is not limited to the type of feature vector used.

The time series of the feature vectors extracted by the feature vector extraction unit 102 is sent to the pattern matching unit 103. A predetermined recognition candidate (recognition unit) is stored in the standard feature pattern storage unit 104 as an HMM having a mixed multidimensional normal distribution learned in advance.

The pattern matching unit 103 performs pattern matching between the time series of feature vectors sent from the feature vector extraction unit 102 and the HMM having the mixed multidimensional normal distribution recorded in the standard feature pattern storage unit 104. To calculate the matching score.

In this embodiment, the pattern matching unit 1
03 calculates the output probability based on the following equation (2). Pattern matching is performed using the value of the output probability.

[0024]

[Equation 2]

However, in the equation (2), d (y-μ)
Represents the distance between the vectors y and μ.

FIG. 4 shows a normal distribution of output probabilities in the pattern matching. Here, for simplicity, the number of dimensions of the normal distribution is 1.

FIG. 5 is a diagram for explaining the function and effect of this embodiment. In the figure, y represents a feature vector at a certain time. The output probability for the feature vector y is P for the learned normal distribution, but in the case of the present embodiment, it is Pc according to the above equation (2).

That is, in the present embodiment, even if the output probability distribution of the speaker acoustically different from the speaker prepared as the learning data is represented by the broken line, the learned normal distribution Within the distance D from the average vector,
The value of the output probability is always the same value Pc. Therefore, due to the deviation of the normal distribution curve, the output probability value becomes unreasonably low at the steep portion, that is, the original value on the normal distribution curve indicated by the wavy line is the point P on the normal distribution curve based on the learning data. It can be avoided that the value becomes a value, and as a result, the matching score does not decrease and the recognition accuracy does not decrease.

According to the first embodiment of the present invention, even when the HMM having the mixed multidimensional normal distribution is learned with the limited learning data, the collation score is lowered for the speaker who has not learned. High-accuracy voice recognition can be performed without performing the above.

[Second Embodiment] Next, a second embodiment of the present invention will be described. The second embodiment relates to the modification of the first embodiment described above, and the basic configuration of the present embodiment is the same as that shown in FIG. 1 of the first embodiment.

The first embodiment described above is characterized in that the output probability is calculated based on the equation (2). The second embodiment is configured such that, when calculating output probabilities, a range that gives equal output probabilities is set independently for each dimension of the feature vector (dimension of mixed multidimensional normal distribution).

In this case, the output probability can be calculated as in equation (3).

[0033]

[Equation 3]

FIG. 6 is a diagram for explaining the function and effect of the second embodiment. In FIG. 5 described in the first embodiment, for simplification, it is assumed that the dimensions of the feature vector are uncorrelated, that is, the covariance matrix is diagonal.

FIG. 6 shows a first-dimensional normal distribution having a variance σi and a mean μi, and a second-dimensional normal distribution having a variance σj and a mean μj. According to the second embodiment, based on the above equations (4) and (5), it is possible to change the range that gives equal output probabilities for each dimension i, j. Therefore, it is possible to obtain the same operational effect as that of the first embodiment described above, and it is possible to adjust the robustness of the output probability in consideration of the magnitude of the variance of the learned normal distribution.

The present invention is not limited to the above-described embodiment, but can be implemented with various modifications.

For example, in the above embodiment, the speech recognition using the HMM having the mixed multidimensional normal distribution has been described, but the HMM having the multidimensional normal distribution instead of the mixture,
The present invention is also applicable to speech recognition using an HMM having a normal distribution that is neither mixed nor multidimensional.

[0038]

As described above, according to the present invention,
In matching likelihood calculation in speech recognition using an HMM having a normal distribution, equal output probabilities are used for input vectors in a predetermined range centered on the average vector of such a normal distribution, so that they are prepared as learning data. It is possible to avoid deterioration of the matching likelihood for a speaker having various acoustic characteristics that deviates from the range of the utterance. Therefore, high-accuracy voice recognition is possible, and a great effect is obtained in practical use.

[Brief description of drawings]

FIG. 1 is a block diagram showing a basic configuration of a voice recognition device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing a configuration of a hidden Markov model (HMM).

FIG. 3 is a diagram for explaining a problem of deterioration of recognition accuracy in conventional speech recognition, and is a diagram for explaining comparison of output probabilities calculated from two different normal distributions on a graph.

FIG. 4 is a diagram showing an output probability distribution when equal output probabilities are given to feature vectors within a predetermined range from the average vector.

FIG. 5 is a diagram for explaining an operation effect according to the first embodiment of the present invention.

FIG. 6 is a diagram for explaining the function and effect according to the second embodiment of the present invention, showing an output probability distribution in the case where a range giving equal output probabilities is changed according to the variance of the normal distribution. Figure

[Explanation of symbols]

101 ... Vocal section detecting unit 102 ... Feature vector extraction unit 103 ... Pattern matching unit 104 ... Standard pattern storage unit 105 ... Recognition result determination unit

Claims (6)

[Claims] 1. A feature vector sequence is extracted by detecting a voice utterance section of an input voice signal and analyzing a voice signal for each of the detected utterance periods, and extracting the feature vector sequence and a predetermined feature vector sequence. Prepared in advance for each recognition candidate,
In a voice recognition method for calculating a matching score by performing pattern matching with a hidden Markov model having a normal distribution, and determining a recognition result of the voice based on the calculated matching score, the pattern matching is the hidden Markov model. A likelihood calculation of the feature vector series based on, and the likelihood calculation gives equal output probabilities to the feature vector series within a predetermined distance range from the average vector of the hidden Markov model. Voice recognition method. 2. A feature vector series is extracted by detecting a voice utterance section from an input voice signal and analyzing the detected voice signal for each utterance section, and extracting the feature vector series and a predetermined feature vector series. Prepared in advance for each recognition candidate,
In a voice recognition method for calculating a matching score by performing pattern matching with a hidden Markov model having a mixed multidimensional normal distribution, in a voice recognition method for determining a recognition result of the voice based on the calculated matching score, the pattern matching is the Including likelihood calculation of the feature vector series based on the hidden Markov model, and the likelihood calculation provides equal output probabilities to the feature vector series within a predetermined distance range from the average vector of the hidden Markov model. Speech recognition method characterized by. 3. The speech recognition method according to claim 2, wherein a range that gives equal output probabilities to the feature vector series is set independently for each dimension of the mixed multidimensional normal distribution. 4. A voicing section detecting means for detecting a voicing section of a voice from an input voice signal, and a feature vector extraction for extracting a feature vector series by analyzing the voice signal of the voicing section detected by the voicing section detecting means. Means, storage means for storing a hidden Markov model prepared for each predetermined recognition candidate and having a normal distribution, feature vector series extracted by the feature vector extraction means, and hidden Markov model stored in the storage means Pattern matching means for calculating the matching score for each recognition candidate by performing the pattern matching, and outputting the feature vector series so as to give an equal value within a predetermined range from the average vector of the hidden Markov model. Pattern matching means including a calculation means for calculating the probability, and each recognition candidate obtained by the pattern matching means Speech recognition apparatus comprising a determining recognition result determining means for collating the recognition result based on the score of the. 5. A voicing section detecting means for detecting a voicing section of a voice from an input voice signal, and a feature vector for extracting a feature vector sequence by voice analysis of the voice signal of the voicing section detected by the voicing section detecting means. Extraction means, storage means that is prepared for each predetermined recognition candidate, and stores a hidden Markov model having a mixed multidimensional normal distribution, feature vector series extracted by the feature vector extraction means, and the storage means. A pattern matching means for calculating a matching score for each recognition candidate by performing pattern matching with a hidden Markov model, wherein the feature is provided so as to give equal values within a predetermined range from the average vector of the hidden Markov model. A pattern matching means including a calculation means for calculating the output probability of the vector sequence; Speech recognition apparatus comprising a determining recognition result determination means results recognized on the basis of the matching score of each recognition candidate. 6. A voicing section detecting means for detecting a voicing section of a voice from an input voice signal, a feature vector sequence is extracted by analyzing a voice signal of the voicing section detected by the voicing section detecting means. Vector extraction means, storage means for storing a hidden Markov model prepared for each predetermined recognition candidate and having a normal distribution, feature vector series extracted by the feature vector extraction means, and hidden Markov models stored in the storage means Pattern matching means for calculating the matching score for each recognition candidate by performing the pattern matching, and outputting the feature vector series so as to give an equal value within a predetermined range from the average vector of the hidden Markov model. Pattern matching means including a calculation means for calculating the probability, each obtained by the pattern matching means Speech recognition program to function as determining the recognition result determining unit a recognition result based on the matching score of each identified candidate.