JP3207378B2 - Voice recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method, and more particularly, to a technique for efficiently searching for a hypothesis closest to an input speech from many hypotheses that can be generated by a given grammar.

[0002]

2. Description of the Related Art First, an example of a conventional speech recognition method will be described with reference to FIG. In this figure, an input speech 11 is converted by an analysis processing unit 12 into a time series of feature parameter vector data.
Is compared with a hypothesis permitted by the grammar / language model 16. The score that is the evaluation value of the matching result of this hypothesis is composed of an acoustic model corresponding to the hypothesis, an acoustic score indicating the likelihood of the input speech, and a language score corresponding to the probability that the hypothesis exists. This is output as the recognition result 14.

[0003] A linear prediction analysis (LPC: Linear Prediction Analysis) is often used as a signal processing in the analysis processing 12.
dictive coding), and the feature parameter is L
There are PC cepstrum, LPC delta cepstrum, mel cepstrum, log power and the like. As the acoustic model 15, a Hidden Markov Model (hereinafter referred to as HMM) modeled based on probability / statistics theory is used.
Is called the law). Details of the HMM method are disclosed in, for example, a document (edited by the Institute of Electronics, Information and Communication Engineers, written by Seiichi Nakagawa, “Speech Recognition by Stochastic Model”).

[0004] The grammar / language model 16 defines the connection of words for defining a sentence to be recognized, and uses a word network with words as branches, a language probability model, or the like. The probability model of this language uses the existence probability of a single word and the probability that two or more words are connected. The details of the stochastic model of this language are disclosed in, for example, literature (edited by the Institute of Electronics, Information and Communication Engineers, written by Seiichi Nakagawa, “Speech recognition using a stochastic model”).

[0005] The search processing unit 13 collates an acoustic model corresponding to a word sequence on a word network indicating a connection relationship of words defined by a grammar with a time series of vector data of characteristic parameters, and indicates acoustic likelihood. While obtaining an acoustic score, a language score is obtained from a language model corresponding to the word sequence, a score of a hypothesis composed of the acoustic score and the language score is obtained at each time, a hypothesis with a low score is discarded, and a hypothesis with a high score is left, At the next time, if necessary, word expansion is performed on the hypothesis left at the previous time, and the evaluation is performed again based on the acoustic model and the language model.

Next, the details of this collation calculation will be described with reference to FIG.
This will be described with reference to FIG. A word network indicating a connection relationship between words specified by a grammar is used for matching with an acoustic model.
The network is deployed in the M state as shown in FIG. In FIG. 2, reference numeral 21 denotes a state indicating a word boundary.

FIG. 3 is a partially enlarged view of FIG. 2, in which reference numerals 31, 32, and 33 indicate states, respectively.
Reference numerals 5 and 36 indicate transitions of the respective states. In each of the states 31, 32, and 33, the score of the hypothesis that the voice from the start time to the current time corresponds to the state from the start state is calculated and stored as follows. Note that the transition 36 indicates the stay of the state 32.

Here, to explain the collation calculation, the calculation of the state 32 at a certain time t will be described. State 3
The hypotheses at the time (t-1) in the states 31, 32, and 33, which are the states of the transition source of No. 2, are transitions 34, 35, and 3, respectively.
Then, the state shifts to the state 32 via FIG. At this time, the acoustic score of the transition 34 with respect to the feature parameter vector data at time t is added to the hypothesis that has proceeded through the transition 34. The hypotheses that transitioned from transitions 35 and 36 are also
The sound score corresponding to No. 6 is added. Of the hypotheses that have reached state 32, the hypothesis with the highest score is the hypothesis in state 32 at time t. The calculation in the state 32 at the time t is similarly performed for all the other states, and the calculation is performed for all the times of the input voice for each time.

At time 0 (initial time), a hypothesis with a score of 0 (zero) is placed at the start of the state network for initialization. As a result of this collation calculation, the hypothesis that has reached the end state of the state network at the end time is the recognition result, and the past word sequence and score information of the hypothesis become the word sequence and score of the recognition candidate. This calculation is
It is called bi-computation, for example, in the literature (edited by the Institute of Electronics, Information and Communication Engineers, edited by Seiichi Nakagawa, Speech Recognition by Stochastic Model)
Is disclosed. However, this collation calculation is performed for all states at each time of the input voice, but if the number of words to be recognized increases or the grammar increases, the total number of states in the state network increases, and the amount of calculation increases. Becomes enormous and recognition requires a lot of processing time.

In order to increase the processing time, a method called beam search is often used. In this beam search, only the hypothesis having a high score is left after the completion of the state matching calculation at each time, and the other hypotheses are rejected. Then, in the collation calculation at the next time, only the hypothesis left at the previous time is calculated. In this case, since it is not necessary to perform calculations for all states, recognition can be performed in a realistic processing time. As a criterion to leave a hypothesis in such a beam search, a method of leaving a maximum of n hypotheses in all the hypotheses, or a score obtained by subtracting a constant value θ from a maximum score in all the hypotheses as a threshold, There is a method of leaving only a hypothesis having a score equal to or greater than the threshold.
Here, n is a parameter for determining a criterion for keeping a hypothesis.
And θ are called beam widths.

In this beam search, generally, when the beam width is reduced, the number of calculation states is reduced, and the recognition processing time is shortened. However, when the beam width is made too small, the recognition accuracy is reduced. In speech recognition using such a beam search, setting of a beam width sets a trade-off between recognition accuracy and recognition processing amount.
The beam width is empirically set depending on the difficulty of recognition, such as the environment in which the recognition is performed, the number of words to be targeted, and the like. Usually, a constant value is used at all times of the input voice.

[0012]

In general, when using a speech recognition apparatus, speech units such as words, phrases, sentences, etc. must be determined, and the boundaries of each speech unit must be indicated to the speech recognition apparatus. . As a method for the speaker to explicitly give the boundary, a press talk method in which a voice can be input only while the voice input key is being used, a method in which voice input is started when the voice input key is pressed, and the like are used. . Although these methods can almost certainly convey a speech section to a speech recognition device through speaker training, they cannot be used in all applications.

In order to improve the convenience by reducing the burden on the speaker, automatic detection of a voice section using a power value or the like at each time is used. It is difficult to eliminate them, and if a detection error occurs, a speech section is lost, resulting in a large decrease in recognition accuracy.

Further, when speech recognition is performed on a continuous word or a sentence, there is a silent section between words or phrases. In the case of recognizing continuous words and sentences collectively, recognition processing equivalent to that of a speech section is necessary for this silent section.
In order to avoid this, there is a method of performing voice recognition by dividing words and phrases, etc., but the speaker inputs the breaks of words, phrases, etc. Not only is it inconvenient, but also requires training because words and phrases must be separated and uttered.

The present invention has been made in view of the above-mentioned problems, and has the following objects. (1) To provide a speech recognition method capable of reducing the burden on a speaker and improving convenience. (2) To provide a speech recognition method capable of reducing the amount of recognition processing in a silent section.

[0016]

In order to achieve the above object, the present invention provides, as a first means, a word list defining a set of words as a unit of recognition, and a word list included in the word list. An acoustic model for obtaining an acoustic score indicating the closeness between the input voice and the input voice, a grammar or language model that defines a connection relationship between words included in the word list, and words permitted by the grammar for the input voice. Means for calculating the likelihood of one or more hypotheses regarding the series as a score at each time based on the acoustic model and the language model, and at the time when all voices are input or during input, the highest score is obtained. In a speech recognition method in which one or a plurality of hypotheses are used as a recognition result, a means for controlling the way of narrowing down the hypotheses according to the likelihood of silence in the input speech is employed. As a second means, in the first means, the method of detecting the likelihood of silence includes a method of narrowing down a hypothesis according to a time during which the power value of each time of the input voice continues to be higher than the set threshold value. A means of controlling is adopted. As a third means, in the first means, the method of detecting the likelihood of silence depends on the time during which the state of the score of the hypothesis matching with silence is higher than the score of the hypothesis matching with non-silence continues. Means to control the hypothesis narrowing.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, an embodiment of a speech recognition method according to the present invention will be described with reference to the drawings. In this embodiment, the beam search is used.
It is characterized in that the beam width is narrowed down in a silent section. In this case, as a method of beam search, either a method of leaving a fixed number of hypotheses as described above or a method of leaving a hypothesis equal to or larger than a threshold obtained from the maximum score can be applied. I will explain how to leave the hypothesis.

First, the silence likelihood P _s (t) at each time is obtained by one or a combination of the two methods described below. The silence likeness P _s (t) takes a value in the range of 0 to 1, and 0
Indicates a voice section (that is, a pause section), and 1 indicates a silent section. As a method for detecting the silent section, for example, the following two methods can be considered.

(1) Method based on power value for each frame When a level higher than the noise level measured at the start of recognition by L (dB) is set as a threshold value and the absolute power value at each time is lower than the threshold value, Calculates the silence likeness P _s (t) according to the duration time. For example, pre-set the maximum duration t _p, the silence likelihood P _s (t) if the duration has a value of 0 to t _p / 2 and as "0", and duration is t _p / silence likeness P in the case of 2~t _p
Assign _s (t) linearly to values from 0 to 1.

(2) Method of Using Hypothesis Having Maximum Score In this method, the hypothesis of the maximum score in the state network is checked at each time, and the state where the hypothesis is present is the HMM state corresponding to silence. Occasionally, silence likeness P _s (t) is obtained according to the duration time. For example, set in advance t _p the maximum duration, assigned the silence likelihood P _s (t) to "0" when the duration time length of 0 to t _p / 2, also duration is t _p / 2 assign silence likelihood P _s (t) linearly to the value of 0 to 1 in the case of ~t _p.

As described above, the silence at each time point P _s
(T) is determined, and a case where the silence likeness P _s (t) is “0” is determined to be a voice section, and the beam width is set to the maximum beam width n.
_max . On the other hand, the case where the likelihood of silence P _s (t) is “1” is determined to be a silent section, and the beam width is set to the minimum beam width n.
_{Make min} correspond linearly. Then, a beam search is performed using the beam width set corresponding to the silence likelihood P _s (t). By thus narrowing the beam width in the silent section, the amount of recognition processing in the silent section can be reduced.

Next, a description will be given of the results of a recognition experiment in the case where the above technique is used in a large vocabulary continuous speech recognition system using a word graph developed by the present inventors. The large vocabulary continuous speech recognition system is described in the literature (Noda et al .: "Approximate Calculation Method in Large Vocabulary Continuous Speech Recognition Using Word Graph", Proceedings of the Acoustical Society of Japan, 2-6-6.
Mar. 1997).

In this experiment, the sentences of newspaper articles were recognized. Symbols were deleted from newspaper articles, and the units divided based on morphological analysis were defined as words. There are 589 and 493 types of words according to such a definition, and of these, the most frequently occurring 9212 words were used as the vocabulary. This vocabulary is
It accounts for 90% of the words that appear in newspaper articles. Also,
Since there are homonyms, the number of readings was 7,703. A word bi-gram corresponding to this vocabulary was learned using 55 months of learning text. This word bi-gram is
Includes those considering the beginning and end of sentences.

The speech data for evaluation selects a sentence composed of only a defined vocabulary set of 9212 words, that is, a sentence without unknown words, out of the text for evaluation for 5 months. Thirty sentences (four men and two women) with relatively many pause sections were selected and used from the audio data recorded (including air conditioning sounds and computer fan sounds). The ratio of the pause time length to the total voice data time length in these 30 sentences was 22.8% on average.

As for the parameters in the method for detecting each silent section, in the method using the power value for each frame, L = 2.0 (dB) and t _p = 0.288 (se
c) In the method using the HMM state having the largest score,
Using t _p = 0.192 (sec), the minimum beam width is 10
0 and the maximum beam width were 5000 and 10000.

Tables 1 and 2 show the experimental results under such conditions. Table 1 shows that the maximum beam width is 5000
Table 2 shows the experimental results when the maximum beam width is 10,000. In these Tables 1 and 2, “power” and “score” indicate each of the above silent section detection methods, and “known” indicates that the pause information is known (data labeled in the pause section is used for search processing). During the search, the minimum beam width is set in the pause section.) “TOP-1” and “WG” indicate the first word recognition rate and the maximum word recognition rate included in the word graph, respectively, and “WGD” indicates the word graph density. The "processing amount" is obtained by normalizing the CPU time in the computer.

[0027]

[Table 1]

[Table 2]

From these Tables 1 and 2, it is confirmed that there is no significant difference between the two methods of detecting silent sections, and that although the recognition accuracy is slightly reduced, the processing amount can be reduced stably. Was completed. Note that the voice data used for the evaluation this time has many portions that are uttered continuously,
Although the overall reduction in processing volume was small,
The ratio of the pause time length to the total audio data length is 2 on average
Considering that it is 2.8%, it can be understood that the processing amount in the pause section can be reduced to about half by such a silent section detection method.

As described above, a word or a phrase is used as a speech unit,
We propose a method to control the beam width of the search based on pose information as an intermediate method between continuous speech recognition, which recognizes sentences while separating them and batch recognition of sentences. went. When two methods were tried for the detection of the pause section and the control method of the beam width, the processing amount of the pause section could be reduced to about half, although the recognition accuracy was slightly lowered.

[0030]

As described above, the speech recognition method according to the present invention has the following effects. (1) Since the amount of recognition processing in a silent section can be reduced, the load on the CPU is not applied until a voice is input. As a result, it is possible to always operate speech recognition in a real system, and it is not necessary to detect the start and end of the speech deliberately. Since it is possible to avoid this and the speaker does not need to explicitly specify the start and end of the voice, the convenience for the speaker is improved. (2) Also, when performing speech recognition for a continuous word or sentence, the processing amount can be automatically suppressed in a silent section, so that a speaker can input a break of a word or a phrase, It is not necessary for the speaker to utter words or phrases in accordance with the automatic detection of the voice section, and the convenience is improved. The present invention is an intermediate method between a method of performing recognition while separating words and phrases as speech units and a method of performing continuous speech recognition in which sentences are collectively recognized. As described above, both advantages are simultaneously obtained. Can be

[Brief description of the drawings]

FIG. 1 is a block diagram showing an outline of a conventional speech recognition process.

FIG. 2 is a state network diagram in the voice recognition processing shown in FIG. 1;

FIG. 3 shows Vite in the voice recognition process shown in FIG.
FIG. 4 is a state network diagram for explaining rbi calculation.

[Explanation of symbols]

 11 input voice 12 analysis processing 13 search processing 14 recognition result 15 acoustic model 16 grammar / language model 21, 31, 32, 33 state 34, 35, 36 transition

Continuation of the front page (56) References JP-A-4-84197 (JP, A) JP-A-8-115093 (JP, A) JP-A-6-167993 (JP, A) Proceedings of the Acoustical Society of Japan VOL. 1997, Fall 1 PAGE 89-90 "Study of Dynamic Search Control Method Based on Pose Information" (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 15/18

Claims (3)

(57) [Claims] 1. A word list defining a set of words as a unit of recognition, an acoustic model for obtaining an acoustic score indicating a closeness between words included in the word list and input speech, and the word list The grammar or language model that defines the connection relationship between the words contained in the and the likelihood of one or more hypotheses about the word sequence allowed by the grammar for the input speech,
Means for calculating a score at each time based on the acoustic model and the language model, and at or at the time of input of all voices, a voice having one or more hypotheses with the highest score as a recognition result. A speech recognition method, wherein a method of narrowing down hypotheses is controlled according to silence in input speech. 2. The speech recognition method according to claim 1, wherein
As a method for detecting the likelihood of silence, a method of narrowing down hypotheses is controlled according to a time during which a power value of an input voice at each time is higher than a set threshold. 3. The speech recognition method according to claim 1, wherein
As a method for detecting the likelihood of silence, it is characterized by controlling the method of narrowing down the hypothesis according to the time during which the state of the score of the hypothesis that is collated with silence is higher than the score of the hypothesis that is collated with non-silence continues. Voice recognition method to be used.