JPH09311693A - Speech recognition apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To lessen the quantity of voice data in learning (registering words) by installing a referring and judging part to compute the likelihood of a voice dictionary file and of each word model and judge a recognition candidate, and a judgment result sending out part. SOLUTION: Voice containing word voice is sent out in a voice input means 1-1 and only a word voice part is cut out of the voice in a word voice cutting out part 1-2. A characteristic extracting part 1-3 extracts the characteristic data from the cut out word voice and a state number estimating part 1-4 estimates the state number corresponding to a word voice at the time of modeling by a hidden Markov model(HMM) from the obtained characteristic data. A learning part 1-5 computes a HMM parameter by matching the characteristic data to the word model. A voice dictionary file 1-6 comprises the learned HMM parameter and the likelihood information and a referring and judging part 1-7 judges a recognized candidate by computing the likelihood for each word model and a judgment result sending out part 1-8 sends out the recognition result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for recognizing a word voice of a specific speaker and outputting the recognition result.

[0002]

2. Description of the Related Art The conventional Hidden Markov
First, in order to explain a voice recognition device for recognizing a word voice using Model, Hidden Marko
A method of voice recognition by vModel will be described.

Hidden Markov Model
(Hereinafter abbreviated as HMM) is the N states S ₁ ,
S ₂ ,. . . , S _N , and transition the states one after another with a certain probability (transition probability) at regular intervals, and at that time,
It is a Markov model that outputs labels (feature data) one by one with a certain probability (output probability). When the speech is regarded as a time series of labels (feature data), each word is uttered several times during learning to create an HMM that models them, and at the time of recognition, the label series of the input speech is output. Recognition is performed by searching for an HMM having the maximum probability (likelihood).

A detailed description will be given below with reference to the drawings.
FIG. 5 is an example diagram of an HMM, which is “Hidden Markov M”, Vol. 42, No. 12 (1986) of the Acoustical Society of Japan.
This is a simple example of an HMM shown in "Old-based speech recognition", and this HMM is composed of three states and outputs a label sequence consisting of only two types of labels a and b.
The initial state is S ₁ , and transition from S ₁ to S ₁ itself with a probability of 0.3 (at that time, label a is output. Since label b has an output probability of 0.0, it is not output), or 0 Transition to S ₂ with a probability of 0.7 (in this case, label a is output with a probability of 0.5 and label b is output with a probability of 0.5).

From the state S _{2, the} state transits to S ₂ itself with a probability of 0.2 (at that time, whether the label a or b is 0.3,
Output with a probability of 0.7) or transition to the final state S ₃ with a probability of 0.8 (in this case, the label b is output. Since the output probability of the label a is 0.0, it is not output). ing. Here, considering the probability (likelihood) that this HMM outputs a label sequence (sequence of feature data) abb, the state sequences allowed by this HMM are S ₁ S ₁ S ₂ S ₃ and S ₁ S ₂ S ₂
There are only two, S ₃ , and the probabilities are 0.3 * 1.0 * 0.7 * 0.5 * 0.8 * 1.0 = 0.0840 0.7 * 0.5 * 0. 2 * 0.7 * 0.8 * 1.0 = 0.0392. Both possibilities are possible, so total 0.0840+
With a probability of 0.0392 = 0.1232, this HMM has ab
It can be seen that b is output.

Therefore, if the HMM for each word is learned in advance and the transition probability of the state most suitable for each word and the output probability of the label at each state transition are obtained, the label sequence of an unknown word is input. In such a case, if the probability (likelihood) is calculated for each HMM, it is possible to know which word the HMM is likely to output this label sequence, and it is possible to recognize it. The above is the method of voice recognition by the HMM.

[0007] FIG. 6 is an example diagram showing the correspondence between the voice waveform, the time series of the characteristic data, and the respective states of the HMM, and shows the correspondence when "beginning" is uttered. in this way,
The HMM is expressed in a state in which the number of phonemes of the word is small with respect to the time series of the voice feature data.

In a conventional speech recognition apparatus for recognizing a word speech using an HMM, at the time of learning, for each word registered in the speech recognition apparatus, the number of states having a small number of phonemes of the word is determined from a change in the phoneme spectrum or the like. Obtained, the output probability of the feature data in each state transition and the transition probability between states are estimated by learning and modeled in HMM, and at the time of recognition, the input speech is applied to all these models to perform likelihood calculation,
I was aware. In addition, it is generally said that a large amount of voice data is required for learning HMMs, but the Acoustical Society of Japan Proceedings (March 3-6-17, March 1991) “Voice Dial Recognition Method for Car Phones As described in ", the variance is used independently for the recognition vocabulary, which is calculated from the voice data of the unspecified speaker and is used in common for each HMM." H is the only voice
MM can be learned.

[0009]

When a speech recognition device returns a recognition result in real time at the time of recognition, all the input speech is input in order to find a word portion from the speech including the input word speech. After the end, instead of determining the word part by obtaining the time series of the feature data, the feature data is actually obtained at each time when the input voice comes in (every 32 ms, which is also called analysis frame synchronization). To determine whether there is a word part. In this case, the recognition candidates are calculated for each time, but narrowing down the candidates is important in order to appropriately reject the candidates that may cause erroneous recognition.

Regarding this, as described in "Spotting of Telephone Voice by HMM" in Technical Research Report of the Institute of Electronics, Information and Communication Engineers (1990, SP90-18), the duration of the state of the HMM (what is the same state? It is necessary to perform processing for narrowing down candidates such as duration control, etc. considering whether or not it should be repeated twice. In the above report, for each word, the standard deviation of each regression coefficient and duration is calculated by assuming a quadratic regression model between the word length and the duration of the HMM state, and finally In addition, it is said that the judgment of the reject is made based on whether or not the predicted value of the state duration and the actual duration are within a certain allowable range.

As described above, at present, statistical information regarding the transition of the HMM state is obtained for each word at the time of learning,
At the time of recognition, the information is weighted. However, in this method, unlike the learning of the HMM itself in which a value obtained in advance for each word can be used, since it is necessary to obtain statistical information for each word, a large amount of voice data, that is, utterance, is required during learning. However, the burden on the user increases. Therefore, there is a problem in that it is not easy to use.

Therefore, the present invention uses Hi for recognizing word speech.
In a voice recognition device using the dden Markov model, statistical information for each word is not used for information necessary for narrowing down recognition candidates, and the number of voice data (utterances) at the time of learning (word registration) is set. It is an object of the present invention to provide a voice recognition device capable of reducing the number.

[0013]

The feature data used for the recognition using the HMM is a cepstrum which is the result of the inverse Fourier transform mainly performed by regarding the logarithm of the power spectrum of the original waveform as the waveform. This cepstrum contains two different pieces of information, the vocal tract and the sound source, and usually the learning / recognition is performed using only the vocal tract information included in the low-order component. By the way, according to the principle and structure of Chapter 12, 12.2 Speech recognition, written by Sozo Saito and Kazuo Nakata, "Basics of Speech Information Processing" (Ohmsha), "The fundamental frequency of the driving sound source characteristic obtained by acoustic analysis is Information such as the sound source amplitude, etc. These are used for preprocessing for roughly classifying phonemes, such as the range of voice intervals and the distinction between voiced and unvoiced sounds, and are used as standard patterns. It is said that there is nothing.

The voice recognition device of the present invention includes a voice input means for inputting a voice including a word voice, a word voice cutout portion for cutting out only a portion of the word voice from the voice including the word voice, and a cutout word voice. A feature extraction unit that extracts feature data, a state number estimation unit that estimates the number of states for a word voice when modeling with HMM from the feature data, a learning unit that applies the feature data to a word model and obtains HMM parameters, A voice dictionary file including the learned HMM parameters, a matching determination unit that calculates a likelihood for each word model and determines a recognition candidate, and a determination result output unit that outputs a recognition result are provided.

With this configuration, at the time of learning, an HMM (to be referred to as a vocal tract HMM) is first obtained using vocal tract information (low-order components of the cepstrum). Furthermore, H prepared separately by using the information of the sound source (higher-order components of the cepstrum)
MM (to be referred to as a sound source HMM) is obtained. At the time of recognition, first, the vocal tract HMM obtained using the information of the vocal tract is used for recognition to obtain a recognition candidate. Next, the sound source HMM obtained by using the sound source information is used to recognize these recognition candidates, and the candidates are narrowed down.

[0016]

According to the present invention, since it is not necessary to obtain statistical information for each word, it is possible to learn by uttering about once or two times at the time of learning and to narrow down candidates at the time of recognition.

A speech recognition apparatus according to an embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a voice recognition apparatus according to an embodiment of the present invention. 1-1 is a voice input means for inputting a voice including a word voice, and 1-2 is a word voice. A word voice cutout unit that cuts out only a word voice portion from a voice, 1-3 is a feature extraction unit that extracts feature data from the cut out word voice, and 1-4 is a state for the word voice when modeling the feature data by HMM. A state number estimating unit for estimating the number, 1-5 is a learning unit for applying the feature data to the word model to obtain HMM parameters, 1
-6 is a voice dictionary file composed of learned HMM parameters and likelihood information, 1-7 is a collation determination unit that calculates a likelihood for each word model and determines a recognition candidate, 1-8
Is a determination result output unit that outputs a recognition result.

FIG. 2 is a circuit block diagram of a voice recognition device according to one embodiment of the present invention.
-2 is a read-only memory (ROM), 2-3 is a central processing unit (CPU), 2-4 is a writable memory (RA)
M), 2-5 is a monitor, and 2-6 is a file device.

The voice input means 1-1 shown in FIG. 1 is realized by a microphone 2-1. Word voice cutout unit 1-
2, feature extraction unit 1-3, state number estimation unit 1-4, and learning unit 1
-5 The collation determination unit 1-7 is realized as a program stored in the ROM 2-2. Voice dictionary file 1-
6 is a file device 2-6, and a determination result output unit 1-8
Are realized by the monitor 2-5.

FIG. 3 is a flowchart for registration in one embodiment of the present invention, and FIG. 4 is a flowchart for recognition in one embodiment of the present invention. In the voice recognition device according to the embodiment of the present invention configured as described above,
In the case where a certain word voice is registered, this operation will be described below with reference to the flowchart of FIG.

In step (3-1), the voice input means 1
With -1, the voiced voice including the word voice is input. In step (3-2), the word voice cutout unit 1-2 cuts out the word voice from the voiced voice including the word voice. This can be realized by detecting and removing silent or low noise portions before and after the word voice by the power of the voice or the like. In step (3-3), the feature extraction unit 1-3 performs feature extraction by obtaining a cepstrum for the word voice by cepstrum analysis. Step (3-
In 4), the state number estimation unit 1-4 causes the step (3-
The number of states for the word voice is estimated from the feature data extracted from the word voice in 3). For this, two types are obtained, one using a low-order component that is vocal tract information and the other using a high-order component that is sound source information. The number of states can be estimated by
This can be done based on the Acoustical Society of Japan Proceedings (1990.3) "On the number of HMM states and the number of mixtures in continuous digit speech recognition".

In step (3-5), the learning unit 1-5 learns the feature data of the word voice using the HMM model having the number of states obtained in step (3-4), and the transition probability between the states. Then, an HMM composed of the output probabilities of the feature data in the transition is obtained, and the obtained HMM is stored in the voice dictionary file 1-6. This is a vocal tract HMM when a low-order component that is vocal tract information is used as the feature data, and a sound source HMM when a high-order component that is a sound source information is used.
Ask for the type.

Next, in the case of recognizing a certain word voice, this operation will be described below with reference to the flowchart of FIG. In step (4-1), the voice input means 1-1
Thus, the uttered voice including the word voice is input. In step (4-2), the feature extraction unit 1-3 performs feature extraction on the input voice. In step (4-3), the matching determination unit 1-7 performs likelihood calculation on the vocal tract HMM of each word model read from the voice dictionary file 1-6 using the low-order component of the input voice feature data. Some word models with high likelihood are determined as recognition candidates. For example, four words (a1, a2, a3, a4) (not shown) from the highest likelihood are used as recognition candidates, and their likelihoods are 0.
It is assumed that they are 98, 0.97, 0.95, 0.87 (not shown). For convenience of the following description, these likelihoods will be referred to as “vocal tract likelihoods”.

In step (4-4), the collation judging unit 1-
Step (4-3) of reading from the voice dictionary file 1-6 by using the higher-order component of the feature data of the input voice according to 7.
The likelihood is calculated on the sound source HMMs of several word models as recognition candidates, and the final judgment is performed by narrowing down to those with high likelihood. For example, the results of likelihood calculation with these sound source HMMs for the four words a1, a2, a3, and a4 shown in step (4-3) are 0.54, respectively.
It is assumed that they are 0.98, 0.98, and 0.88. When these likelihoods are called “sound source likelihoods”, recognition candidates can be narrowed down as follows. Likelihood for narrowing down candidates = α × (likelihood of vocal tract) + β × (likelihood of sound source), where α + β = 1.

These coefficients can be predetermined by evaluating the device. For example, if this example is calculated with α = 0.8 and β = 0.2, the likelihood of narrowing down the candidates of the word a1 is 0.8 × 0.98 + 0.2 × 0.54 = 0.89.
2 Likelihood of narrowing down candidates for word a2 is 0.8 × 0.97 +
0.2 × 0.98 = 0.972 The likelihood of narrowing down candidates for the word a3 is 0.8 × 0.95 +
0.2 × 0.98 = 0.956 The likelihood of narrowing down the candidates of the word a4 is 0.8 × 0.87 +
The result is 0.2 × 0.88 = 0.872, and the word a2 having the highest likelihood finally becomes the recognition result. In step (4-5), the determination result output unit 1-8 notifies the user of the recognition result.

[0027]

According to the speech recognition apparatus of the present invention, since it is not necessary to obtain statistical information for each word even when narrowing down candidates, learning can be performed by uttering once or twice. When a person registers (learns) a word, there is no inconvenience such as having to speak a number of times, and therefore the usability for the user can be improved.

[Brief description of drawings]

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

FIG. 2 is a circuit block diagram of a voice recognition device according to an embodiment of the present invention.

FIG. 3 is a flowchart at the time of registration according to the embodiment of the present invention.

FIG. 4 is a flowchart at the time of recognition according to the embodiment of the present invention.

FIG. 5 is an example diagram of Hidden Marcov Model.

FIG. 6 is an example diagram showing a correspondence between a voice waveform, a time series of characteristic data, and HMM states.

[Explanation of symbols]

 1-1 Voice input means 1-2 Word voice cutout unit 1-3 Feature extraction unit 1-4 State number estimation unit 1-5 Learning unit 1-6 Voice dictionary file 1-7 Collation determination unit 1-8 Determination result output unit 2-1 Microphone 2-2 Read-only memory (ROM) 2-3 Central processing unit (CPU) 2-4 Writable memory (RAM) 2-5 Monitor 2-6 File device

Claims (1)

[Claims] 1. A voice input unit for inputting a voice including a word voice, a word voice cutout unit for cutting out only a portion of the word voice from the voice including the word voice, and a feature for extracting feature data from the cut out word voice. It consists of an extraction unit, a state number estimation unit that estimates the number of states for word speech when modeling with the HMM from the feature data, a learning unit that applies the feature data to the word model to obtain HMM parameters, and a learned HMM parameter. A voice dictionary file,
A speech recognition apparatus comprising: a matching determination unit that calculates a likelihood for each word model and determines a recognition candidate; and a determination result output unit that outputs a recognition result.