JPH05257491A - Voice recognizing system - Google Patents

Abstract

PURPOSE:To obtain a system for recognizing a voice at a high speed without dropping accuracy. CONSTITUTION:An input voice is analyzed by a voice analyzing part 1, an analyzed feature parameter is applied to a conversion part 3 and a symbol sequence is found out by using a symbol recognizing dictionary 2. The symbol sequence is passed through the 1st HMM recognizing part 5 in which the 1st HMM is set up, and probability of outputting of the symbol sequence by the model is found out. Plural words are extracted by an upper candidate extracting part 6 in the order of decending probability. Then, the symbol sequence is passed through the 2nd HMM recognizing part 8 in which the 2nd HMM having more states than the 1st HMM is set up and probability of outputting of the symbol sequence by the model is found out. A word is specified based upon the probability of the recognized results.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition system for recognizing spoken voice.

[0002]

2. Description of the Related Art A method of performing vector quantization for converting speech into a constant symbol sequence and recognizing the quantized symbol sequence by a hidden Markov model (hereinafter referred to as HMM) has been successful in recent years.

A general formulation of the HMM will be described. H
It is assumed that the MM has n states S1, S2, ..., Sn, and the initial state is stochastically distributed to these n states. For speech, a model that transitions a state with a certain probability (transition probability) for each fixed frame period is used. At the time of a transition, a null transition that transitions the state without outputting a symbol with a certain probability (output probability) may be introduced. Even if the output symbol sequence is given, the state transition sequence is not uniquely determined. Since only the symbol series can be observed, it is called a hidden Markov model. The HMM model M is defined by the following six parameters. n: Number of states (states S1, S2, ..., Sn) h: Number of symbols (symbols R1, R2, ..., Rh) Pij: Probability of transition probability Si to Sj Qij (h): From Si to Sj Probability of outputting symbol h during transition mi: initial state probability (probability that initial state is Si) F: set of final states

Next, the model M is subject to transitional restrictions that reflect the characteristics of speech. In speech, a loop transition that returns from a state Si to a previously passed state (Si-1, Si-2, ...) Is not allowed because it disturbs the temporal context. As a structure of the HMM described above, an example as shown in FIG. 5 is typical. The HMM is evaluated by the probability Pr (O) that the model M outputs the symbol series O = o1, o2, ..., Ot.
/ M). At the time of recognition, the HMM recognition unit assumes each model and obtains a model M that maximizes Pr (O / M) by the Viterbi algorithm. Also, for learning HMM, the HMM learning unit gives O to a large number of symbol sequences,
It suffices to estimate the parameter of the model M that maximizes Pr (O / M) on average. By recognizing the input voice uttered as described above, the input voice can be recognized with high accuracy.

[0005]

According to the above-mentioned conventional technique, when the number of words in the calculation process for recognition reaches about 200, the current workstation is in real time (within 200 to 300 ms after utterance). In (), there is a problem that the calculation amount becomes large so that the result cannot be obtained. The present invention has been made in view of the above circumstances, and an object thereof is to provide a voice recognition system for performing recognition processing at high speed without degrading accuracy.

[0006]

In order to solve the above problems, the present invention provides a voice analysis means for inputting a voice signal and performing voice analysis to obtain a feature parameter, and a feature parameter obtained by the voice analysis means. Conversion means for converting into a symbol series, and a first probability determining means for passing this symbol series through a first hidden Markov model created in advance for each word, and obtaining a probability that the model outputs the symbol series, Based on the probabilities obtained by the first probability determining means, the upper candidate extracting means for extracting a plurality of words from the one with the highest probability and the plurality of words extracted by the upper candidate extracting means A second pre-created second with more states than the first Hidden Markov Model
Pass this symbol sequence to the hidden Markov model of and obtain the probability that the model outputs this symbol sequence.
The probability determining means is provided and the word is specified based on the probability determined by the second probability determining means to perform voice recognition.

Further, according to the present invention, the first probability determining means is configured as a set of k stages (k is an integer of 2 or more), and a first preliminarily prepared for each set of n words. The first probability determining means for passing the symbol sequence through a plurality of models among the hidden Markov models of the above, and determining the probability that the model outputs this symbol sequence, and the probability determined by the first probability determining means. Initially, a high-rank candidate extraction means for extracting m words (m is an integer satisfying 2 ≦ m <n) from the highest probability is provided. Then, the first probability determining means in the first stage passes the symbol sequence to all of the first hidden Markov models for each of the n words, and the first probability determining means other than the first stage uses the preceding stage combination. The first hidden Markov model for the m words extracted by the higher-rank candidate extracting means is passed through the symbol sequence. The k-stage set is configured such that the number of states of the first hidden Markov model is increased in the subsequent stage, and the number m of words extracted by the upper candidate extracting means is decreased in the subsequent stage. And this k
The symbol sequence is applied to a second hidden Markov model created in advance, which has a larger number of states than the first hidden Markov model for the m words extracted by the higher-rank candidate extraction means in the final stage of the stage set. And a second probability determining means for determining the probability that the model outputs this symbol sequence, and performing speech recognition by specifying a word based on the probability determined by the second probability determining means. Is characterized by.

[0008]

According to the above configuration, a voice signal is input and voice analysis is performed to obtain a characteristic parameter, the characteristic parameter is converted into a symbol series, and the symbol series is generated in advance for each word. Through the hidden Markov model,
The first probability determining means determines the probability that the model outputs the symbol sequence. Based on the probabilities obtained by the first probability determining means, a plurality of words with the highest probability are extracted by the high-rank candidate extracting means, and first, rough recognition is performed as the first step to determine the words of the input speech. Give a candidate.

For the plurality of words extracted by the upper candidate extraction means, the symbol sequence is passed through a second hidden Markov model created in advance, which has a larger number of states than the first hidden Markov model, and the model thereof is passed. The second probability determining means determines the probability that the symbol output the symbol sequence, and the word is specified based on the determined probability.

In this way, the accuracy is lowered by performing detailed recognition by the second hidden Markov model having a large number of states only for the word candidates given by the rough recognition in the first stage. The recognition process is performed at high speed.

Further, not only the recognition in two stages as described above, but at the beginning, the recognition is roughly performed to extract word candidates, and the candidates are recognized in more detail to narrow down the candidates. It is also possible to perform recognition in multiple stages.

[0012]

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

(First Embodiment) FIG. 1 is a block diagram of a voice recognition apparatus showing a first embodiment of the present invention. Figure 1
The speech recognition device includes a speech analysis unit 1, a symbol recognition dictionary 2, a conversion unit 3, a first HMM setting unit 4, and a first HMM.
The recognition unit 5, the high-rank candidate extraction unit 6, the second HMM setting unit 7, and the second HMM recognition unit 8 are provided.

The voice analysis unit 1 analyzes the input voice and extracts characteristic parameters. The symbol recognition dictionary 2 is an identification dictionary created from a plurality of standard patterns for each symbol. The conversion unit 3 performs a matching process between the characteristic parameter extracted by the voice analysis unit 1 and a predetermined symbol registered in the symbol recognition dictionary 2 to obtain a symbol series.

The first HMM setting section 4 prepares an HMM (first HMM) for each of, for example, 32 words prepared in advance.
MM) is set in the first HMM recognition unit 5. First H
The MM recognizing unit 5 receives the symbol sequence obtained by the converting unit 3 and performs a process for obtaining the probability that each of the set first HMMs outputs this symbol sequence. The higher-rank candidate extraction unit 6 extracts a plurality of words from the one with the highest probability of this processing result.

The second HMM setting section 7 prepares an HMM (second HMM) for each of, for example, 32 words prepared in advance.
MM) is set in the second HMM recognition unit 8. Second H
The MM recognition unit 8 outputs the probability that each HMM outputs this symbol sequence through the symbol sequence obtained by the conversion unit 3 to the above-set second HMM for the plurality of words extracted by the high-rank candidate extraction unit 6. And the word is specified based on this probability.

FIG. 2 shows the first HMM set in the first HMM recognition unit 5 of FIG. This HMM is left
To-right type, five states S1, S2, ...,
It is a model that has S5, has only S1 as the initial state, and outputs a symbol with a certain output probability in a frame period of 8 ms. The parameters for, for example, 32 models of the HMM of this system are as follows. n: number of states = 5 (states S1, S2, ..., S5) h: number of symbols = 191 (R = 1, 2, ..., 191 which code each symbol) Pij: transition probability S i to S j Probability Qij (h): Probability of outputting the symbol h at the transition from Si to Sj Further, the final probability is limited to S5.

FIG. 3 shows a second HMM set in the second HMM recognizing unit 8 in FIG. This HMM is left
10 states of to right type S1, S2, ...,
It has S10, the initial state is only S1, and the state is transited with a constant transition probability in a frame period of 8 ms. It is a model that outputs a symbol with a constant output probability at the time of the transition. The parameters for the 32 HMM models of the system in this example are as follows: n: number of states = 10 (states S1, S2, ..., S10) h: number of symbols = 191 (R = 1, 2, ..., 191 which code each of the symbols) Pij: transition probability S i to S j Probability Qij (h): Probability of outputting the symbol h at the transition from Si to Sj Further, the final probability is limited to S10. Next, the voice recognition processing with the configuration of FIG. 1 will be described.

When a voice is input, the voice analysis unit 1 analyzes it by, for example, a linear prediction method (LPC) or a bandpass filter (BPF) and extracts a characteristic parameter. The conversion unit 3 performs a matching process on the analyzed and extracted characteristic parameters with the standard pattern for each symbol registered in the symbol recognition dictionary 2 to obtain a symbol sequence.

The first HMM setting unit 4 learns and accumulates the first HMM shown in FIG. 2 for predetermined 32 words in advance, and sets this HMM in the first HMM recognition unit 5. The symbol sequence obtained for the input speech is passed through the first HMM in the first HMM recognition unit 5 and the probability Pr of outputting it through the upper candidate extraction unit 6
(O / M) is calculated, and an HMM is calculated so that this probability is from the 1st place to the 5th place.

The second HMM setting unit 7 preliminarily learns and accumulates the second HMM for a predetermined 32 words having a larger number of states than the first HMM shown in FIG. 2 is set in the HMM recognition unit 8. Second H
In the MM recognition unit 8, the above-mentioned symbol sequence is passed through the second HMM of the five words from the first place to the fifth place obtained by the high-rank candidate extraction unit 6, and the probability Pr (O / O /
M) is obtained, and the word having the highest probability is used as the recognition result.

In the recognition process, as the number of states of the HMM increases, the accuracy of the recognition result obtained by obtaining the probability through the HMM symbol sequence increases, but the time required for the recognition process increases. become longer. When the number of states is small, the accuracy of the recognition result is low, but the time required for the recognition process is short.

Therefore, in this embodiment, the recognition process is performed in two steps, and in the first step, the symbol sequence of the input voice is passed through the HMM having the number of states of 5 for 32 words, and the probability of outputting this is obtained, 5 Hs from the highest probability
Find MM.

Next, in the second step, an HMM having a number of states of 10 for 32 words is prepared, and the symbol sequence of the input voice is converted into an HMM of five words from the one having the highest probability obtained in the first step. Throughout, the word with the highest probability is used as the recognition result.

To give a concrete example of the above-mentioned recognition processing time, for example, in the conventional system in which the recognition processing is not performed in two stages,
If the recognition result is the word that has the maximum probability of performing recognition processing for the 32 words with the HMM state number 10, it is about 60 m.
Processing is performed at s.

On the other hand, in the present embodiment in which the recognition process is performed in two stages, first, the process of the first stage, that is, the recognition process is performed with the number of HMM states being 5, and the five words with the highest probability are selected. Is extracted in about 30 ms. Next, the process of the second stage, that is, the process of recognizing the five words extracted in the process of the first stage with the HMM state number as 10 and recognizing the word with the highest probability as the recognition result, It is performed in 10 ms. Therefore, according to this embodiment, the time is reduced by about 1/3 as compared with the conventional method. Moreover, as a result of the test, no final error was increased.

(Second Embodiment) In the first embodiment, the first HMM recognition unit 5 and the second HMM recognition unit 8 are used.
The word was specified by performing the processing in two stages by using the two probability determining means of the above. However, as in the second embodiment described below, 3
It is also possible to specify a word by performing processing in steps or more.

FIG. 4 is a block diagram of a voice recognition apparatus showing the second embodiment, and the same parts as those in FIG. 1 are designated by the same reference numerals. The voice recognition device of FIG. 4 includes a voice analysis unit 1, a symbol recognition dictionary 2, a conversion unit 3, a second HMM setting unit 7, and a second HMM recognition unit 8 as in the voice recognition device of FIG. The first HMM setting unit 4 of FIG. 1, the first
Instead of the HMM recognizing unit 5 and the higher-rank candidate extracting unit 6, the k-stage first HMM setting unit 4-i connected in series multi-stage, the first
HMM recognition unit 5-i and higher-rank candidate extraction means 6-i (i =
1-k).

The configuration will be described only with respect to this difference.
The first HMM setting unit 4-i is the first HMM recognition unit 5-i.
For example, H for each of 32 words prepared in advance
Set MM (first HMM). First-stage 1st HM
The M recognizing unit 5-1 receives the symbol sequence obtained by the converting unit 3 and obtains the probability that each first HMM for each of 32 words outputs this symbol sequence. The upper candidate extraction unit 6-1 extracts m1 words (m1 is an integer satisfying 2≤m1 <32) from the one with the highest probability of the processing result of the first HMM recognition unit 5-1. The HMM recognition unit 5-j (j = 2 to k) in the second and subsequent stages is the upper candidate extraction unit 6- (j-1) in the previous stage.
The probability that each first HMM for m (j-1) words extracted by is output this symbol sequence is obtained. Then, the higher-rank candidate extraction unit 6-j receives the first HMM recognition unit 5-j.
Mj pieces (mj is 2 ≦ m
The words of j <32 and mj <m (j-1) are extracted. Here, the number of states of the first HMM is set to increase in the subsequent stages. Next, the voice recognition processing with the configuration of FIG. 4 will be described. The processes up to the voice analysis unit 1, the symbol recognition dictionary 2, and the conversion unit 3 are the same as those in the first embodiment.

The first HM setting unit 4-1 has a first HM.
M is learned in advance for 32 predetermined words, and this H
The MM is set in the first HMM recognition unit 5-1. The symbol sequence obtained by the conversion unit 3 is converted into the first HMM recognition unit 5
At -1, the probability of outputting this symbol sequence is obtained by passing through this first HMM. The high-rank candidate extraction unit 6-1 obtains the probability of outputting this, and extracts m1 words (m1 is an integer satisfying 2 <m1 <32) from the highest probability.

First HMM setting unit 4-j in the second and subsequent stages
(J is an integer of 2≤j≤k), the first HM of the preceding stage
The HMM for a predetermined 32 words having a larger number of states than the HMM of the M setting unit 4- (j-1) is learned in advance, and this HMM is set in the first HMM recognition unit 5-j. Then, in the HMM recognition unit 5-j, the upper-rank candidate extraction unit 6 in the preceding stage
The probability that the above first HMM outputs the symbol sequence for m (j-1) words extracted by-(j-1) is obtained. The high-rank candidate extraction unit 6-j extracts mj words from the one with the highest probability of the processing result. Here, since the number mj of words extracted by the higher-rank candidate extraction unit 6-j is set to be smaller in the subsequent stage, the candidate words are gradually narrowed down. Further, since the number of states of the first HMM is set to be larger in the latter stage, the recognition process can be performed in more detail in the latter stage for the candidate word. At this time, the number of candidate words decreases as the latter stage as described above, and therefore the processing time can be reduced even if the number of HMM states increases as the latter stage.

The second HMM setting unit 7 preliminarily learns and accumulates the second HMM for a predetermined 32 words having a larger number of states than the first HMM, and recognizes this HMM as the second HMM recognition. Set in part 8. In the second HMM recognition unit 8, the above-mentioned symbol sequence is passed through the second HMM for the mk words obtained by the k-th (final stage) higher-rank candidate extraction unit 6-k and is output. The probability is calculated, and the word having the highest probability is used as the recognition result. Although the discrete HMM has been described in the above embodiment, the continuous HMM is used.
The same is performed by the MM.

[0033]

As described above in detail, according to the present invention,
First, a hidden Markov model with a small number of states (first hidden Markov model) is used to obtain the probability (first probability) of outputting the symbol sequence obtained from the feature parameter of the input speech to be recognized, A large classification is performed by extracting a plurality of words (extracting upper candidates) from the one with a higher probability and roughly selecting the candidate, and then a hidden Markov model having a larger number of states than the hidden Markov model (second
Hidden Markov model), the probability of outputting a symbol sequence (second probability) is calculated for a plurality of words from the one with the highest extracted probability, and the word of the input speech is identified based on this probability. With this configuration, the recognition processing can be performed at high speed without lowering the accuracy.

Further, according to the present invention, in obtaining the above-mentioned first probability, the first hidden Markov model having a larger number of states in the subsequent stages is provided in multiple stages, and the first hidden Markov model in the preceding stage is provided in each stage. The higher the probability of outputting the symbol sequence to the multiple words (the number of words is smaller than that extracted in the previous stage, but in the first stage all input speech words) from the one with the highest probability of being extracted by With the configuration obtained in stages, when the number of recognition target words is large, the recognition processing can be performed at a higher speed without lowering the accuracy.

[Brief description of drawings]

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

FIG. 2 is an H set in the first HMM recognition unit 5 of FIG.
The figure which shows MM.

3 is an H set in a second HMM recognition unit 8 in FIG.
The figure which shows MM.

FIG. 4 is a block diagram showing a second embodiment of a voice recognition device to which the invention is applied.

FIG. 5 is a diagram showing a general HMM.

[Explanation of Codes] 1 ... Speech analysis unit, 2 ... Symbol recognition dictionary, 3 ... Conversion unit,
4, 4-1 to 4-k ... First HMM setting unit, 5, 5-1 to 5
-k ... 1st HMM recognition part (1st probability determination means), 6,
6-1 to 6-k ... High-rank candidate extraction unit, 7 ... Second HMM setting unit, 8 ... Second HMM recognition unit (second probability determining means).

Claims (2)

[Claims] 1. A voice analysis unit for inputting a voice signal and performing voice analysis to obtain a characteristic parameter, a conversion unit for converting the characteristic parameter obtained by the voice analysis unit into a symbol sequence, and the symbol sequence as a word. A first hidden Markov model that is created in advance for each model, and a first probability determining unit that obtains a probability that the model outputs the symbol sequence; and a probability that is obtained by the first probability determining unit. In addition, a higher-rank candidate extraction means for extracting a plurality of words from the one having a higher probability, and a plurality of states of the plurality of words extracted by the higher-rank candidate extraction means, which have more states than the first hidden Markov model, are created in advance. Second probability determining means that passes the symbol sequence through a second hidden Markov model and obtains the probability that the model outputs the symbol sequence Comprising a speech recognition system, characterized in that the probability determined by the second probability determination unit configured to identify words based. 2. A voice analysis means for inputting a voice signal and performing voice analysis to obtain a characteristic parameter, a conversion means for converting the characteristic parameter obtained by the voice analysis means into a symbol sequence, and k stages (k is 2). The above integer sequence), each set passes the symbol sequence through a plurality of models of the first hidden Markov model created in advance for each of the n words, and the model is the symbol sequence. Based on the probabilities obtained by the first probability determining means and the probability obtained by the first probability determining means, and m probabilities from the larger probability (m is an integer satisfying 2 ≦ m <n ), The first probability determining means at the first stage has a first candidate for each of the n words.
Pass the above symbol sequence to all hidden Markov models of
The first probability determining means other than the first stage is configured to pass the symbol sequence through the first hidden Markov model for the m words extracted by the upper candidate extracting means in the preceding set, A k-stage set configured such that the number of states of the first hidden Markov model increases as the number of states increases, and the number m of words extracted by the higher-rank candidate extraction means decreases as the number of states decreases. The second hidden Markov model having a larger number of states than the first hidden Markov model for the m words extracted by the upper candidate extraction means in the final stage of Second probability determining means for determining the probability that the model outputs the symbol sequence, and based on the probability determined by the second probability determining means. Speech recognition method being characterized in that so as to identify.