JP2001075596A - Voice recognition device, voice recognition method and recording medium -> stored with voice recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a voice recognition device whose recognition accuracy is high without increasing computational cost. SOLUTION: A likelihood calculating means 15 calculates hypothetic likelihood by acoustic featured vectors from an acoustic analyzing means 12, acoustical models of respective phonemes and acoustical models of words. A simple acoustical model probability calculating means 22 calculates simple acoustic output probability from the acoustical models of the respective phonemes and the acoustic featured vectors and an order fluctuation calculating means 23 calculates the average of order fluctuation widths of the simple acoustical output probability and a beam width changing means 24 changes a beam width by the average of the order fluctuation widths. A pruning means 16 rejects hypotheses having acoustical likelihood equal to or smaller than the beam width set by the beam width changing means 24 from the maximum likelihood.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to a speech recognition device,
The present invention relates to a voice recognition method and a recording medium storing a voice recognition program.

[0002]

2. Description of the Related Art FIG. 26 is a diagram showing the configuration of a conventional speech recognition apparatus analogized based on Japanese Patent Publication No. Hei 4-22276 and Japanese Patent Laid-Open Publication No. H8-248984. In the figure, 1
Reference numeral 1 denotes an audio data storage unit which digitizes an input audio and stores it as audio data, and 12 denotes an audio which takes in the audio data stored in the audio data storage unit 11 at predetermined times, performs an acoustic analysis, and outputs an acoustic feature vector. An analysis unit 13 is an acoustic model storage unit that stores an acoustic model of each phoneme, and 14 is a word dictionary storage unit that generates and stores an acoustic model of a word from a phoneme description of each word in the word dictionary.

[0006] In FIG. 26, reference numeral 15 denotes an acoustic feature vector output from the acoustic analysis means 12, an acoustic model of each phoneme stored in the acoustic model storage means 13, and stored in the word dictionary storage means 14. A likelihood calculating means 16 for calculating the likelihood of a hypothesis that is a recognition candidate based on the acoustic model of the word, finds the maximum likelihood of the hypothesis, and is within a certain value (hereinafter, referred to as a beam width) from the found maximum likelihood. Is a pruning means for rejecting a hypothesis smaller than the beam width while leaving the hypothesis, and a recognition result output means 17 for outputting recognition words in descending order of likelihood. Acoustic model storage means 13, word dictionary storage means 1
4. The frame recognition processing means 18 is constituted by the likelihood calculating means 15 and the pruning means 16.

FIG. 27 is a flowchart showing the processing of a conventional speech recognition apparatus, FIG. 28 is a diagram showing a specific example of a word dictionary, and FIG. 29 is a diagram showing a structural example of an acoustic model of a word. Numeral 30 is a diagram for explaining a situation in which a hypothesis is developed as the time (hereinafter, also referred to as a frame) advances.

Next, the operation will be described. First, the acoustic model storage unit 13 reads from an external storage device (not shown)
The acoustic model of each phoneme is read and stored. Hereinafter, for the sake of simplicity, the acoustic model will be described as an HMM acoustic model (Hidden Markov Model) of each phoneme, but the same applies to other acoustic models.

Next, the word dictionary storage means 14 reads and stores a word dictionary from an external storage device. The word dictionary, as shown in FIG.
It has Hiragana and Phoneme data. Further, the word dictionary storage unit 14 generates an HMM acoustic model of the word from the phoneme description of each word. FIG. 29 is a diagram illustrating a configuration example of the HMM acoustic model of the word “hashi”. As shown in FIG. 29, each word is decomposed into each phoneme, an HMM acoustic model corresponding to each phoneme is applied, and the resultant is connected to form an acoustic model of the HMM of each word.

Next, the audio data storage means 11 performs A / D conversion of the input audio, converts it into digital data, and stores it as audio data.

In step ST1 of FIG. 27, the control means (not shown) sets the time t to t = 0, and in step ST2, the acoustic analysis means 12 sets the time t to time t (t = 0).
From the voice data storage means 11,
In step ST3, the acoustic analysis unit 12 acoustically analyzes the acquired audio data at t = 0, and calculates an acoustic feature vector.

In step ST4, likelihood calculating means 1
5 calculates the output probability of each HMM state from the acoustic feature vector calculated in step ST3 and the acoustic model of each phoneme stored in the acoustic model storage unit 13.

In step ST5, likelihood calculating means 1
5 reads out a hypothesis that is a candidate for recognition of each word from the word dictionary storage means 14 and calculates each H calculated in step ST4.
From the output probability of the MM state and the HMM acoustic model of the word stored in the word dictionary storage means 14, the acoustic likelihood of each hypothesis is calculated and the hypothesis is developed. A method for calculating the acoustic likelihood is well known, and is disclosed, for example, in Japanese Patent Publication No. Hei 4-22276. Here, the acoustic likelihood is a measure indicating how close the voice data is to the sound.

FIG. 30 is a diagram showing how a hypothesis is developed at each time (frame). For simplicity, the number of HMM states of each phoneme is set to one. In the figure, each square is a hypothesis, and each hypothesis has a recognized word, a current phoneme, and acoustic likelihood as information. Assuming that the recognition word of the hypothesis is "article", as the frame progresses, the HMM acoustic model self-loops and the phoneme does not progress, and the hypothesis that the HMM acoustic model progresses and the phoneme progresses is developed. It will increase. In particular, as shown by the thick framed hypothesis in FIG. 30, when the phoneme / i / at the end of the "article" ends, the transition to the next recognized word occurs.
After "article", "ga", "school", "bridge" ...
, So that various words can transition, hypotheses are assigned to each independently.

With respect to the hypothesis immediately after the word transition in step ST5, the likelihood calculating means 15 may add the connection probability between words to the likelihood calculation in the form of a sum or a product. The probability of connection between words is represented by a conditional probability, and a bigram model, a trigram model, and the like are well known. When the connection probability between words is added to the likelihood calculation, the likelihood calculating means 15 reads the connection between the currently recognized word and the word immediately before or further from the language model storage means (not shown). The probability (language likelihood) is obtained and added to the acoustic likelihood. This linguistic likelihood indicates the transition probability of another word following a certain word.

In such a case, a method of correcting a language probability based on a generalized Bernoulli trial (IEICE Transactions D-
II, Vol. J81-D-II, No12, pp. 2
703-2711) ", the linguistic likelihood is generally added with a certain weight to the acoustic likelihood, and the weight is further added. Is
As disclosed in the same document, it is common to give an optimum value obtained in a preliminary experiment.

Further, the likelihood calculating means 15 may perform the duration control in step ST5. That is, in a certain hypothesis, a duration that is a time remaining in a certain phoneme is measured, and a penalty is given if the duration is different from the duration data obtained in advance. FIG. 31 is a diagram illustrating an example of the duration time. For example,
When a certain hypothesis is expanded as shown in FIG. 31, / k /
Is measured as one frame, the duration of / i / is measured as two frames, etc. The duration data includes an average duration of / k /, an average duration of / i /, etc., which are calculated in advance during learning, and based on the difference. The penalty is determined.

In this case, if the utterance is uttered earlier than the learning data, a large penalty is given, and as a result, a recognition error may be caused. Therefore, in consideration of the difference between the utterance speed in the learning data and the utterance speed in the recognition, the correction is performed, for example, as described in "Evaluation by Keyword Spotting in Word Reject Method (Journal of the Acoustical Society of Japan, Heisei 10 September, pp. 159-16
0) ". Here, a certain period of the utterance is analyzed, the utterance speed is obtained, and the duration of each phoneme is expressed as a linear function of the utterance speed. When the utterance speed changes, the average duration also changes. .

In step ST6 of FIG. 27, the pruning means 16 obtains the maximum likelihood from each of the hypotheses.
At T7, the pruning unit 16 arranges each hypothesis in the order of acoustic likelihood, and sets a constant value (hereinafter, referred to as a beam width) from the maximum likelihood.
Retain only hypotheses with acoustic likelihoods within and reject hypotheses below that.

In steps ST8 and ST9, the above-described processing in steps ST2 to ST7 is performed until the speech is completed. In step ST10, the recognition result output means 17 determines the likelihood of the hypothesis for which the calculation has been completed in all the sections of the utterance. Recognized words (strings) are output as recognition results in descending order of degree.

Further, the above-described processing is performed in step ST9 when t
Is replaced with t + 1 to perform processing at each time (for each frame), but processing may be omitted depending on the frame in order to speed up the processing time. In this case, as described in “Study of Feature Vector Interpolation and Frame Decimation Method in Speech Recognition (Spring 1999 IEICE National Convention, D-14-21)”, processing is skipped every few frames. It is common.

[0019]

Since the conventional speech recognition apparatus is configured as described above, there is a method of increasing the beam width in step ST7 as one of means for improving the recognition performance. However, when the beam width is widened, the possibility that the correct hypothesis is pruned is reduced, but the number of hypotheses remaining in the beam width increases, so that the processing amount increases. In particular, when it is known that the correct answer candidate is in the upper position in the beam, widening the beam width merely causes an increase in calculation cost, and there is a problem that there is much waste.

In the process of step ST5, since the weight of the linguistic likelihood with respect to the acoustic likelihood is fixed, there is a problem that the recognition accuracy is deteriorated depending on the ambiguity of the acoustic.

Further, in the duration control in step ST5, the determination of the speech rate is
There is a problem that it is necessary to analyze a certain utterance section or more, and it takes a long processing time.

Further, regarding the omission of the frame processing in the processing of step ST9, since the presence or absence of the thinning and the ratio (thinning rate) are fixed, the sound is ambiguous and precise likelihood is required. In such a case, there is a problem that a recognition error may be caused.

Further, in the process of step ST5, when a sentence is uttered, a characteristic that a space is often left slightly behind a particle such as "ga" is not reflected. There was a problem that could not be done.

Further, in the processing of step ST5, when words are uttered continuously, the feature that the gap between words is likely to be vacant and the case where vacancy is difficult is not reflected. There was a problem that control might not be possible.

Further, there is another problem that when exhalation or the like is applied to the microphone, it is recognized as a voice and may be erroneously recognized.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and a speech recognition apparatus and a speech recognition system having high recognition accuracy without increasing computational cost, suppressing pruning of a correct answer. It is an object to obtain a recording medium on which a recognition method and a voice recognition program are recorded.

Another object of the present invention is to provide a speech recognition apparatus, a speech recognition method, and a recording medium on which a speech recognition program is recorded, having high recognition accuracy even if there is ambiguity in sound.

A further object of the present invention is to obtain a speech recognition apparatus, a speech recognition method, and a recording medium on which a speech recognition program is recorded, having a high recognition accuracy even for a fast utterance or a slow utterance.

Further, the present invention provides a speech recognition apparatus, a speech recognition method, and a recording medium on which a speech recognition program is recorded with high recognition accuracy even for long utterances in which the utterance speed of a sentence or the like varies locally. The purpose is to gain.

A further object of the present invention is to provide a speech recognition apparatus, a speech recognition method and a recording medium on which a speech recognition program is recorded with high recognition accuracy even when exhalation or the like is applied to a microphone.

[0031]

According to the present invention, there is provided a speech recognition apparatus comprising: a speech data storage unit for converting input speech into digital data and storing the same as speech data; An acoustic analysis unit that outputs an acoustic feature vector, an acoustic model storage unit that stores an acoustic model of each phoneme, a word dictionary storage unit that generates and stores an acoustic model of a word from a phoneme description of each word in a word dictionary, It is a recognition candidate based on the acoustic feature vector output from the acoustic analysis means, the acoustic model of each phoneme stored in the acoustic model storage means, and the acoustic model of a word stored in the word dictionary storage means. A likelihood calculating means for calculating the likelihood of the hypothesis, a maximum likelihood is obtained from the likelihood of the hypothesis calculated by the likelihood calculating means, and a maximum likelihood is obtained from the obtained maximum likelihood. A simple acoustic model of each phoneme is stored in a device including a pruning unit that rejects a hypothesis having a beam width equal to or less than a beam width and a recognition result output unit that outputs a hypothesis left by the pruning unit as a recognition candidate. The simple acoustic model storage means, the acoustic feature vector output from the acoustic analysis means, and the simple acoustic model of each phoneme stored in the simple acoustic model storage means, for each of the predetermined times including the current time. A simple acoustic model probability calculating means for calculating a simple acoustic output probability of each HMM state at a time; and a ranking of the simple sound output probabilities of each HMM state at each time obtained by the simple acoustic model probability calculating means, and Rank change calculating means for calculating a rank change width of each HMM state within a predetermined time and calculating an average of the rank change widths of the HMM states; Based on the average rank variation width of rank variation calculation means has calculated, and adjusts the parameters of the speech recognition.

The speech recognition apparatus according to the present invention includes beam width changing means for setting a beam width as a parameter relating to speech recognition, and is adapted to be used when the average of the order variation calculated by the order variation calculating means is smaller than a predetermined value. The beam width changing means sets the beam width small, and if the average of the order variation width is larger than a predetermined value, the beam width changing means sets the beam width large, and the pruning means The hypothesis is rejected based on the beam width set by the width changing means.

The speech recognition apparatus according to the present invention includes language weight changing means for setting the weight of the language likelihood added to the likelihood as a parameter relating to speech recognition, and the rank variation calculated by the rank variation calculation means. When the average of the width is smaller than a predetermined value, the language weight changing means sets the weight of the language likelihood to be small, and when the average of the order variation width is larger than a predetermined value, the language weight changing means sets the language The weight of the likelihood is set large, and the likelihood calculating means calculates the likelihood of the hypothesis based on the weight of the language likelihood set by the language weight changing means.

A speech recognition apparatus according to the present invention sets duration data as a parameter relating to speech recognition, and a duration data storage means for storing speech duration data corresponding to a plurality of speech rates. When the average of the rank fluctuation range calculated by the rank fluctuation calculating means is smaller than a predetermined value, the duration data changing means is stored in the duration data storage means. If you select the duration time data that was uttered late, and the average of the rank fluctuation range is larger than a predetermined value,
The duration data changing means selects the uttered duration data stored in the duration data storage means, and the likelihood calculating means selects the duration selected by the duration data changing means. The likelihood of a hypothesis is calculated based on the time length data.

The speech recognition apparatus according to the present invention includes thinning-out determining means for determining whether or not frame processing, which is recognition processing at each time, is to be thinned out as a parameter relating to voice recognition. When the average of the fluctuation width is smaller than a predetermined value, the thinning-out determining means sets the thinning of the frame processing to ON, and when the average of the rank fluctuation width is larger than the predetermined value, the thinning-out determining means sets the thinning of the frame processing. Is set to none.

The speech recognition apparatus according to the present invention includes a thinning-out rate determining means for determining a thinning-out rate of a frame process as a recognition process at each time, as a parameter relating to the voice recognition. When the average of the fluctuation width is smaller than a predetermined value, the thinning rate determining means sets the thinning rate of the frame processing to be large, and when the average of the rank fluctuation width is larger than the predetermined value, the thinning rate determining means sets the frame processing. The thinning rate is set to be small.

The voice recognition apparatus according to the present invention converts the input voice into digital data and stores it as voice data. The voice data is taken in at predetermined time intervals, and the voice is analyzed to output a voice feature vector. Acoustic analysis means, an acoustic model storage means for storing an acoustic model of each phoneme, a word dictionary storage means for generating and storing an acoustic model of a word from a phoneme description of each word in a word dictionary, and an output from the acoustic analysis means The likelihood of a hypothesis that is a recognition candidate is calculated based on the obtained acoustic feature vector, the acoustic model of each phoneme stored in the acoustic model storage unit, and the acoustic model of a word stored in the word dictionary storage unit. A maximum likelihood is calculated from the likelihood of the hypothesis calculated by the likelihood calculation means, and a predetermined beam width or less is calculated from the calculated maximum likelihood. In the apparatus provided with a pruning means for rejecting a hypothesis and a recognition result output means for outputting a hypothesis left by the pruning means as a recognition candidate, duration data of utterance corresponding to a plurality of utterance speeds is stored. Word length obtaining means for obtaining the word and the position in the word of each hypothesis currently being recognized from the likelihood calculating means, and obtaining the part of speech of the hypothetical word from the word dictionary storing means; A duration length data changing unit that selects duration length data corresponding to the part of speech and the position in the word of the hypothesis word acquired by the word type acquisition unit from the duration length data storage unit; The means calculates the likelihood of the hypothesis based on the duration data selected by the duration data changing means.

The voice recognition apparatus according to the present invention converts voice data into digital data and stores the voice data as voice data. The voice data is taken in at predetermined time intervals, the voice is analyzed, and a voice feature vector is output. Acoustic analysis means, an acoustic model storage means for storing an acoustic model of each phoneme, a word dictionary storage means for generating and storing an acoustic model of a word from a phoneme description of each word in a word dictionary, and an output from the acoustic analysis means The likelihood of a hypothesis that is a recognition candidate is calculated based on the obtained acoustic feature vector, the acoustic model of each phoneme stored in the acoustic model storage unit, and the acoustic model of a word stored in the word dictionary storage unit. A maximum likelihood is calculated from the likelihood of the hypothesis calculated by the likelihood calculation means, and a predetermined beam width or less is calculated from the calculated maximum likelihood. A language model storage means for storing a chain probability of a word, comprising: a pruning means for rejecting a hypothesis; and a recognition result output means for outputting a hypothesis left by the pruning means as a recognition candidate; A duration data storage unit for storing duration data of the utterance corresponding to the word, and a word of each hypothesis currently being recognized and a word recognized before that are obtained from the likelihood calculation unit, and the language model storage unit is obtained. A word chain probability obtaining means for obtaining the chain probability of those words from the word sequence, and a continuation time period data corresponding to the word chain probability obtained by the word chain probability obtaining means, selected from the continuous time data storage means. Time length data changing means, wherein the likelihood calculating means calculates the likelihood of a hypothesis based on the duration data selected by the duration data changing means. It is.

The voice recognition apparatus according to the present invention converts the input voice into digital data and stores it as voice data. The voice data is stored at predetermined time intervals, and the voice is analyzed to output a sound feature vector. Acoustic analysis means, an acoustic model storage means for storing an acoustic model of each phoneme, a word dictionary storage means for generating and storing an acoustic model of a word from a phoneme description of each word in a word dictionary, and an output from the acoustic analysis means The likelihood of a hypothesis that is a recognition candidate is calculated based on the obtained acoustic feature vector, the acoustic model of each phoneme stored in the acoustic model storage unit, and the acoustic model of a word stored in the word dictionary storage unit. A maximum likelihood is calculated from the likelihood of the hypothesis calculated by the likelihood calculation means, and a predetermined beam width or less is calculated from the calculated maximum likelihood. In a device comprising a pruning means for rejecting a hypothesis and a recognition result output means for outputting a hypothesis left by the pruning means as a recognition candidate, obtaining the maximum likelihood of the hypothesis from the pruning means and obtaining It has a maximum likelihood hypothesis adding means for adding a hypothesis having the maximum likelihood as the first utterance and outputting it to the pruning means.

According to the speech recognition method of the present invention, there is provided a first step of taking in speech data at a predetermined time and performing acoustic analysis to output an acoustic feature vector; an acoustic feature vector output in the first step; A second step of calculating a simple acoustic output probability of each HMM state at each time within a predetermined time including the current time by using a simple acoustic model of each phoneme stored in advance;
A third step of obtaining the order of the simplified sound output probabilities in the MM state, calculating the order variation width of each HMM state within a predetermined time including the current time, and calculating the average of the order variation width of the HMM state; A fourth step of setting a predetermined beam width based on the average of the order variation width, the acoustic feature vector output in the first step, an acoustic model of each phoneme and an acoustic model of a word stored in advance. A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate;
A sixth step of obtaining a maximum likelihood from the calculated likelihood of the hypothesis and rejecting a hypothesis having a beam width equal to or smaller than the predetermined beam width set in the fourth step from the obtained maximum likelihood; And a seventh step of outputting the remaining hypotheses as recognition candidates.

According to the speech recognition method of the present invention, a first step of taking in speech data at predetermined time points and performing acoustic analysis to output an acoustic feature vector; an acoustic feature vector output in the first step; A second step of calculating a simple acoustic output probability of each HMM state at each time within a predetermined time including the current time by using a simple acoustic model of each phoneme stored in advance;
A third step of obtaining the order of the simplified sound output probabilities of the MM states, calculating the order change width of each HMM state within a predetermined time including the current time, and calculating the average of the order change widths of all the HMM states; A fourth step of adjusting parameters related to speech recognition based on the average of the order variation width, an acoustic feature vector output in the first step, an acoustic model of each phoneme stored in advance, and an acoustic of a word. A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate using the model and parameters related to speech recognition adjusted in the fourth step, and a maximum likelihood is calculated from the calculated likelihood of the hypothesis. Recognizing speech by providing a sixth step of rejecting a hypothesis having a predetermined beam width or less from the maximum likelihood and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate Is shall.

According to the voice recognition method of the present invention, a first step of taking in voice data at predetermined time points and performing acoustic analysis to output an acoustic feature vector; an acoustic feature vector output in the first step; A second step of calculating an output probability of each HMM state from a pre-stored acoustic model of each phoneme; acquiring a word and a position in the word of each hypothesis currently being recognized; A third step of acquiring the part of speech of the hypothesis word and the duration data corresponding to the part of speech and the position in the word of the hypothesis word acquired in the third step. A fourth step of selecting from time length data, an output probability of each HMM state calculated in the second step, an acoustic model of a word stored in advance, and the fourth step A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the duration data selected in the step, and obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and a predetermined beam from the obtained maximum likelihood. The speech recognition apparatus includes a sixth step of rejecting a hypothesis having a width less than or equal to a width and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate.

According to the speech recognition method of the present invention, a first step of taking in speech data at predetermined time points and performing acoustic analysis to output an acoustic feature vector, and the acoustic feature vector output in the first step and the acoustic feature vector A second step of calculating an output probability of each HMM state from the stored acoustic model of each phoneme; acquiring a word of each hypothesis currently being recognized and a word recognized before that; A third step of obtaining the chain probabilities of those words from the chain probabilities of the above, and the duration time data corresponding to the chain probabilities of the words obtained in the third step are corresponded to the speech rates stored in advance. A fourth step of selecting from the utterance duration data of the utterance, an output probability of each HMM state calculated in the second step, an acoustic model of a word stored in advance, A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the duration data selected in the fourth step, and obtaining a maximum likelihood from the calculated likelihood of the hypothesis. The speech recognition apparatus includes a sixth step of rejecting a hypothesis having a predetermined beam width or less and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate.

According to the speech recognition method of the present invention, a first step of taking in speech data at a predetermined time and performing acoustic analysis to output an acoustic feature vector; an acoustic feature vector output in the first step; A second step of calculating the likelihood of a hypothesis that is a recognition candidate using the acoustic model of each phoneme and the acoustic model of a word stored in advance, and a maximum likelihood calculated from the likelihood of the hypothesis calculated in the second step. A fourth step of obtaining the maximum likelihood obtained in the third step, adding a hypothesis having the obtained maximum likelihood and making the first utterance, and a third step of The voice recognition includes a fifth step of rejecting a hypothesis having a predetermined beam width or less from the maximum likelihood obtained in the above, and a sixth step of outputting the hypothesis left in the fifth step as a recognition candidate. Is shall.

The recording medium on which the speech recognition program according to the present invention is recorded captures the speech data at predetermined time intervals, analyzes the acoustic data, and outputs an acoustic feature vector. A second step of calculating a simple sound output probability of each HMM state at each time within a predetermined time including the current time, based on the obtained sound feature vector and a simple sound model of each phoneme stored in advance; A third step of calculating the order of the simple sound output probabilities of the respective HMM states at each time, calculating the order fluctuation width of the respective HMM states within a predetermined time including the current time, and calculating the average of the order fluctuation width of the HMM states; And a fourth step of setting a predetermined beam width based on the average of the order variation width.
And a fifth step of calculating the likelihood of a hypothesis that is a recognition candidate using the acoustic feature vector output in the first step, the acoustic model of each phoneme and the acoustic model of a word stored in advance. A sixth step of obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and rejecting a hypothesis having a beam width equal to or smaller than the predetermined beam width set in the fourth step from the obtained maximum likelihood; The seventh step of outputting the hypothesis left in the step as a recognition candidate is executed by a computer.

The recording medium on which the speech recognition program according to the present invention is recorded has a first step of taking in speech data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, and the output in the first step. A second step of calculating a simple sound output probability of each HMM state at each time within a predetermined time including the current time by using the sound feature vector and a simple sound model of each phoneme stored in advance; A third order of obtaining the order of the simplified sound output probabilities of the respective HMM states at the time, calculating the order variation of each HMM state within a predetermined time including the current time, and calculating the average of the order variation of all the HMM states; Step, a fourth step of adjusting a parameter related to speech recognition based on the average of the order variation width, and an acoustic feature vector output in the first step. And a fifth step of calculating the likelihood of a hypothesis that is a recognition candidate by using the acoustic model of each phoneme and the acoustic model of the word stored in advance and the parameters related to the speech recognition adjusted in the fourth step. A sixth step of obtaining a maximum likelihood from the calculated likelihood of the hypothesis and rejecting a hypothesis having a predetermined beam width or less from the obtained maximum likelihood;
The seventh step of outputting the hypothesis left in the sixth step as a recognition candidate is executed by a computer.

The recording medium on which the speech recognition program according to the present invention is recorded has the first step of taking in speech data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, and the output in the first step. From the acoustic feature vector and the acoustic model of each phoneme stored in advance, each H
A second step of calculating the output probability of the MM state, a third step of acquiring a word and a position in the word of each hypothesis currently being recognized, and acquiring a part of speech of a previously stored hypothesis word, A fourth step of selecting, from the pre-stored utterance duration data corresponding to the utterance speed, duration time data corresponding to the part of speech and the position in the word of the hypothesis word acquired in the third step; And the output probability of each HMM state calculated in the second step, the acoustic model of the word stored in advance, and the duration data selected in the fourth step, the likelihood of a hypothesis that is a recognition candidate. A fifth step of calculating the degree of likelihood, a sixth step of obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and rejecting a hypothesis of a predetermined beam width or less from the obtained maximum likelihood; Left in 7 for outputting the hypothesis as a recognition candidate
Are executed by a computer.

The recording medium on which the speech recognition program according to the present invention is recorded has the first step of taking in speech data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, and the output in the first step. From the acoustic feature vector and the acoustic model of each phoneme stored in advance, each H
A second step of calculating the output probability of the MM state, acquiring the word of each hypothesis currently being recognized and the previously recognized word, and calculating the chain probability of those words from the chain probability of the word stored in advance. A third step of obtaining
A fourth step of selecting duration time data corresponding to the word chain probability obtained in the step from the utterance duration data corresponding to the utterance speed stored in advance;
The output probability of each HMM state calculated in the second step, the acoustic model of the word stored in advance, and the fourth
A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the duration data selected in the step, and obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and a predetermined beam from the obtained maximum likelihood. A sixth step of rejecting a hypothesis having a width less than or equal to the width and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate are executed by a computer.

The recording medium on which the speech recognition program according to the present invention is recorded has the first step of taking in speech data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, and the output in the first step. A second step of calculating a likelihood of a hypothesis that is a recognition candidate based on the acoustic feature vector, an acoustic model of each phoneme and an acoustic model of a word stored in advance, and the likelihood of the hypothesis calculated in the second step. A third step of obtaining the maximum likelihood from the degrees, a fourth step of obtaining the maximum likelihood obtained in the third step, adding a hypothesis having the obtained maximum likelihood and the first utterance, A fifth step of rejecting a hypothesis of a predetermined beam width or less from the maximum likelihood obtained in the third step;
The sixth step is to cause a computer to execute a sixth step of outputting the hypotheses left in the fifth step as recognition candidates.

[0050]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 1 of the present invention. In the figure, reference numeral 21 denotes a simple acoustic model storage unit for storing a simple acoustic model of each phoneme; 22, an acoustic feature vector from the acoustic analysis unit 12 and a simple acoustic model of each phoneme stored in the simple acoustic model storage unit 21. All or part of H using a simple acoustic model
This is a simple acoustic model probability calculating means for obtaining a simple acoustic output probability at each time within a time sandwiching the current time (current frame) in the MM state.

In FIG. 1, reference numeral 23 denotes the order of the simplified sound output probability at each time of each HMM state i,
The rank change calculating means for calculating the rank change width within the time including the current time t and obtaining the average of the rank change widths for all or some of the HMM states, the rank change width calculated by the rank change calculating means 23 Is a beam width changing unit that sets a beam width based on the average of the above and outputs the beam width to the pruning unit 16. The other configuration is the same as the conventional configuration shown in FIG. In this embodiment, a beam width is set as a parameter related to speech recognition.

Next, the operation will be described. First, the simple acoustic model storage unit 21 stores, from an external storage device (not shown), an acoustic model that is simpler than the acoustic model of each phoneme used in the recognition processing, for example, an acoustic model of the recognition processing is environment-dependent T
In the case of the ripone model, an environment independent Uniphone model is acquired and stored as a simple acoustic model. Such a simple acoustic model is characterized in that it is smaller in number than the acoustic model used in the recognition processing and does not require a calculation cost.

Next, the audio data storage means 11 performs A / D conversion on the input audio, converts it into digital data, and stores it as audio data.

FIG. 2 shows a voice according to the first embodiment of the present invention.
It is a flowchart which shows the process of a recognition device. Steps
In ST1, time t is set to t = 0 as in the conventional case.
Then, in step ST2, the acoustic analysis means 12
T between times₁= T-Δt to t_Two= T + Δt
In between, audio data is captured. However, the time is negative
If it becomes, do not take in. In step ST3,
The acoustic analysis means 12 calculates t ₁= T-Δt to t_Two= T + Δ
The voice data captured at time t is analyzed and the
Calculate the vector.

In step ST20, the simple acoustic model probability calculating means 22, the rank variation calculating means 23, and the beam width changing means 24 are stored in the simple acoustic model storing means 21 with the acoustic feature vector from the acoustic analyzing means 12. The beam width used by the pruning unit 16 is set based on a simple acoustic model of each phoneme.

FIG. 3 is a flowchart showing the processing for setting the beam width in step ST20 of FIG. In step ST <b> 21, the simple acoustic model probability calculation means 22 calculates the sound feature vector from the sound analysis means 12,
Using simple acoustic models for each phoneme are stored in the simple acoustic model storage unit 21, for all or part of the HMM state, simple acoustic output at each time t _x in a time sandwiching the current time (current frame) Find the probability. That is, when a simple sound output probability of the HMM state i at frame time t and P (t, i), for all or part of the HMM states, t ₂ from t ₁ = t-Delta] t sandwiching the current time t
= T + Δt, the simple sound output probability P (t _x ,
Calculate i). Here, for some HMM states,
As a case of calculating the simple sound output probability P (t _x , i), for example, only the hypothesis existing within the beam width may be calculated.

[0057] In step ST22, rank variation calculating means 23, at each time t _x, simple acoustic output probability P
The order of HMM state i when (t _x , i) is arranged in descending order is determined as R1 (t _x , i), and step ST23 is performed.
In, order variation calculation means 23, sandwich the current time t t ₁ = t-Δt from t ₂ = t + time in a Position variation range of Δt R2 (t, i) = max (R1 (t x, i)) - m
Calculate in (R1 (t _x , i)). Where max
And min are taken over t ₁ ≦ t _x ≦ t ₂ . Step S
At T24, the rank change calculation means 23 calculates the average of the rank change width R2 (t, i) for all or some of the HMM states. Here, as a case where the average of the order variation R2 (t, i) for some HMM states is obtained, for example, the average may be obtained only from the higher-order HMM state in the order obtained in step ST22.

In step ST25, the beam width changing unit 24 sets the beam width to be smaller when the average of the order variation obtained by the order variation calculator 23 is smaller than the predetermined value, and the average of the order variation is larger than the predetermined value. If it is large, set the beam width large.

Here, the meaning of step ST25 will be described. 4 and 5 are diagrams showing the output probability rank R1 at each time of the three HMM states. FIG. 4 shows a case where the rank variation R2 (t, i) is large, and FIG. 5 shows a rank variation R2. The case where (t, i) is small is shown.

As shown in FIG. 4, if the average of the order fluctuation width is larger than a predetermined value, there is a possibility that acoustic fluctuation may occur in the vicinity of the predetermined range, so it is appropriate to increase the beam width. . Further, as shown in FIG. 5, when the order variation width is small, the sound is stable, and in the vicinity thereof, the inversion of the acoustic likelihood is also performed in the main recognition process after step ST4 shown in FIG. Is unlikely to occur,
Even if the beam width is reduced, it is unlikely that the correct answer will be pruned.

Steps ST4 to ST6 in FIG.
This is the same as the processing of steps ST4 to ST6 shown in FIG. In step ST7 of FIG. 2, the pruning unit 16 arranges each hypothesis in the order of the acoustic likelihood, and leaves a hypothesis having an acoustic likelihood within the beam width set by the beam width changing unit 24 from the maximum likelihood. Reject hypotheses with acoustic likelihood from the degree to the beam width. Step ST in FIG.
The processes from 8 to ST10 are the same as the processes from steps ST8 to ST10 shown in FIG.

As described above, according to the first embodiment, since the beam width is appropriately set using the simple acoustic model of each phoneme, the calculation cost does not increase so much and the correct answer can be obtained. The effect of suppressing pruning and improving recognition accuracy can be obtained.

Embodiment 2 FIG. 6 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 2 of the present invention. In the figure, reference numeral 31 denotes a weight of the linguistic likelihood based on the average of the order variation width obtained by the order variation calculating means 23. This is a language weight changing means to be set. In this embodiment, the beam width changing unit 24 in FIG. 1 of the first embodiment is replaced with a language weight changing unit 31, and the output of the language weight changing unit 31 is input to the likelihood calculating unit 15. This is for setting the weight of the linguistic likelihood added to the likelihood as a parameter related to speech recognition.

Next, the operation will be described. FIG. 7 is a flowchart showing processing of the voice recognition device according to the second embodiment of the present invention. The processing from step ST1 to ST3 is the same as the processing from step ST1 to ST3 in FIG. 2 of the first embodiment.

In step ST 30, the simple acoustic model probability calculating means 22, the rank variation calculating means 23, and the language weight changing means 31 are stored in the simple acoustic model storing means 21 and the acoustic feature vector from the acoustic analyzing means 12. The weight of the linguistic likelihood is set based on a simple acoustic model of each phoneme.

FIG. 8 is a flowchart showing the process of setting the weight of the language likelihood in step ST30 of FIG. The processing from steps ST31 to ST34 is performed from steps ST21 to ST24 in FIG. 3 of the first embodiment.
The processing is the same as that described above.

In step ST35, if the average of the order variation width obtained by the order variation calculating means 23 is smaller than a predetermined value, the language weight changing means 31 considers the sound likelihood to be acoustically stable. If the weight of language is set smaller than the language likelihood, and the average of the order variation is larger than a predetermined value, it is considered that the sound is unstable. The language weight is set to be large, with the language likelihood being regarded as more important than the likelihood.

The process in step ST4 in FIG. 7 is the same as the process in step ST4 in FIG. 27 in the related art. In step ST36 of FIG. 7, likelihood calculating means 1
5 reads out the hypothesis of each word from the word dictionary storage means 14 and outputs the output probability of each HMM state calculated in step ST4, the HMM acoustic model of the word stored in the word dictionary storage means 14, and the language weight change. Based on the language weight set by the means 31, the acoustic likelihood of each hypothesis is calculated and the hypothesis is developed.

The processing from steps ST6 to ST10 in FIG. 7 is the same as the processing from steps ST6 to ST10 in FIG.
This is the same as the processing up to 10.

As described above, according to the second embodiment, the language weight is appropriately set using the simple acoustic model of each phoneme, so that the calculation cost does not increase so much.
The effect that the recognition accuracy can be improved can be obtained.

Embodiment 3 FIG. 9 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 3 of the present invention. In the drawing, reference numeral 41 denotes a duration data storage unit that stores duration data of an utterance corresponding to the utterance speed; 42 is
Based on the average of the rank fluctuation width obtained by the rank fluctuation calculation means 23, the duration data stored in the duration data storage means 41 is selected, and the duration data to be used is set. Means. In this embodiment, the language weight changing means 3 in FIG.
1 is replaced by a duration data storage unit 41 and a duration data change unit 42, which sets duration data as a parameter related to speech recognition.

Next, the operation will be described. First, the duration time data storage means 41 is an external storage device (not shown).
, And obtains and stores respective duration data corresponding to different utterance speeds such as fast utterance, normal utterance, and slow utterance.

FIG. 10 is a flowchart showing processing of the voice recognition device according to the third embodiment of the present invention. The processing from steps ST1 to ST3 is the same as that of the first embodiment shown in FIG.
Is the same as the processing from step ST1 to ST3.

In step ST 40, the simple acoustic model probability calculating means 22, the rank variation calculating means 23, and the duration data changing means 42 are stored in the simple acoustic model storing means 21 with the acoustic feature vector from the acoustic analyzing means 12. Based on the simple acoustic model of each phoneme, the duration data stored in the duration data storage unit 41 is selected, and the duration data to be used is set.

FIG. 11 is a flowchart showing the process of setting the duration data in step ST40 of FIG. Steps ST41 to ST44 are the same as the processing of steps ST21 to ST24 in FIG. 3 of the first embodiment.

In step ST45, if the average of the order change width calculated by the order change calculating means is smaller than a predetermined value, the duration time data changing means 42 determines that the acoustically stable state has continued for a long time. Judgment and selection of the uttered duration data stored later in the duration data storage means 41, and setting the duration data to be used, when the average of the rank fluctuation width is larger than a predetermined value. Is
When it is determined that the sound is changing quickly, the utterance duration data that is uttered earlier stored in the duration data storage unit 41 is selected, and the duration data to be used is set.

The processing in step ST4 of FIG.
This is the same as the processing in step ST4 in FIG. 27 of the related art. In step ST46 of FIG. 10, the likelihood calculating means 15 reads out the hypothesis of each word from the word dictionary storage means 14, and the HM included in each hypothesis calculated in step ST4 is included.
Based on the output probability of the M state, the HMM acoustic model of the word stored in the word dictionary storage unit 14, and the duration data set by the duration data change unit 42, the acoustic likelihood of each hypothesis is calculated. Expand your hypothesis.

The processing from steps ST6 to ST10 in FIG. 10 is the same as the processing from steps ST6 to ST in FIG.
This is the same as the processing up to T10.

As described above, according to the third embodiment, since the duration data corresponding to the utterance speed is set by using the simple acoustic model of each phoneme, the calculation cost is significantly increased. Instead, even if the speed of the utterance changes, the effect of increasing the recognition accuracy can be obtained.

Embodiment 4 FIG. 12 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 4 of the present invention. In the figure, reference numeral 51 denotes whether frame processing is to be omitted based on the average of the order variation width obtained by the order variation calculating means 23. This is a thinning-out determining means for determining whether or not to perform the determination. In this embodiment, the language weight changing means 31 in FIG. 6 of the second embodiment is replaced with a thinning-out determining means 51, and a frame processing which is a recognition processing at each time as a parameter relating to speech recognition is performed. This is to determine the presence or absence of thinning.

Next, the operation will be described. FIG. 13 is a flowchart showing processing of the voice recognition device according to the fourth embodiment of the present invention. The processing from step ST1 to ST3 is the same as the processing from step ST1 to ST3 in FIG. 2 of the first embodiment.

In step ST 50, the simple acoustic model probability calculating means 22, the order variation calculating means 23, and the thinning determining means 51 determine the acoustic feature vector from the acoustic analyzing means 12 and each phoneme stored in the simple acoustic model storing means 21. It is determined whether to omit the frame processing based on the simple acoustic model.

FIG. 14 is a flowchart showing a process for determining whether or not to omit the frame process in step ST50 in FIG. Steps ST51 to ST5
Step 4 is the same as the processing from step ST21 to ST24 in FIG. 3 of the first embodiment.

In step ST55, the thinning-out determining means 51 is acoustically stable when the average of the rank fluctuation width calculated by the rank fluctuation calculating means 23 is smaller than a predetermined value. If it is determined that the number is small, it is set to be thinned. If the average of the order variation width is larger than a predetermined value, it is determined that it is necessary to calculate accurately, and the thinning is set.

In step ST56 in FIG. 13, the control means (not shown) determines whether or not the current frame is a predetermined frame to be thinned out when the thinning is set in step ST55. It is checked, and if it is a thinning target, the frame processing of steps ST4 to ST7 is omitted, and the process proceeds to step ST8. If it is not a thinning target, the frame processing of steps ST4 to ST7 is executed. Also, when no thinning is set in step ST55, steps ST4 to ST7 are performed.
Execute the frame processing up to.

The processing from steps ST8 to ST10 in FIG. 13 is the same as the processing from steps ST8 to ST in FIG.
This is the same as the processing up to T10.

As described above, according to the fourth embodiment, the use of a simple acoustic model of each phoneme determines whether or not to perform thinning in accordance with the situation of voice data. Therefore, the calculation cost does not increase so much. This has the effect of increasing the recognition accuracy.

Embodiment 5 FIG. 15 is a diagram showing the configuration of a speech recognition apparatus according to Embodiment 5 of the present invention. In the figure, reference numeral 61 denotes a frame processing thinning rate determined on the basis of the average of the order variation width calculated by the order variation calculating means 23. This is the thinning rate determining means. In this embodiment, the thinning-out determining means 51 in FIG. 12 of the fourth embodiment is replaced with a thinning-out rate determining means 61. Determine the rate.

Next, the operation will be described. FIG. 16 is a flowchart showing processing of the voice recognition device according to the fifth embodiment of the present invention. The processing from step ST1 to ST3 is the same as the processing from step ST1 to ST3 in FIG. 2 of the first embodiment.

In step ST60, the simple acoustic model probability calculating means 22, the rank variation calculating means 23, and the thinning rate determining means 61 determine the acoustic feature vector from the acoustic analyzing means 12 and the respective data stored in the simple acoustic model storing means 21. A thinning rate of frame processing is determined based on a simple acoustic model of a phoneme.

FIG. 17 is a flow chart showing the processing for determining the thinning rate of the frame processing in step ST60 of FIG. Steps ST61 to ST64 correspond to steps ST21 to ST2 in FIG. 3 of the first embodiment.
This is the same as the processing up to 4.

In step ST65, the thinning-out rate determining means 61 is acoustically stable when the average of the rank fluctuation width calculated by the rank fluctuation calculating means 23 is smaller than a predetermined value. Judging that deterioration is small,
If the frame processing thinning rate is set to be large and the average of the order variation width is larger than a predetermined value, it is determined that accurate calculation is necessary, and the frame processing thinning rate is set to be small.

In step ST66 in FIG. 16, the control means (not shown) checks whether the current frame is a frame to be thinned out based on a predetermined thinning rate. ST4 to S
The frame processing up to T7 is omitted, and the process proceeds to step ST8.
The frame processing up to T7 is executed.

The processing from steps ST8 to ST10 in FIG. 16 is the same as the processing from steps ST8 to ST in FIG.
This is the same as the processing up to T10.

As described above, according to the fifth embodiment, the thinning rate according to the situation of the voice data is determined using the simple acoustic model of each phoneme, so that the calculation cost does not increase so much. The effect that the recognition accuracy can be improved can be obtained.

Embodiment 6 FIG. FIG. 18 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 6 of the present invention. In the figure, reference numeral 41 denotes duration time data storage means for storing duration time data of utterance corresponding to utterance speed; 71 is a word type obtaining unit that obtains the word of each hypothesis currently being recognized and the position in the word from the likelihood calculating unit 15 and obtains the part of speech of the hypothetical word from the word dictionary storage unit 14. 72 is a word type obtaining unit 71. Is the duration length data changing unit that selects duration length data corresponding to the part of speech and the position within the word of the hypothesis acquired from the duration length data storage unit 41 and sets the duration length data to be used. . In this embodiment, a duration data storage unit 41, a word type acquisition unit 71, and a duration data change unit 72 are added to the conventional configuration shown in FIG.

Next, the operation will be described. First, the duration time data storage means 41 is an external storage device (not shown).
, And obtains and stores respective duration data corresponding to different utterance speeds such as fast utterance, normal utterance, and slow utterance. FIG. 19 is a flowchart showing processing of the voice recognition device according to the sixth embodiment of the present invention. Steps ST1 to ST4 correspond to step ST1 in FIG.
This is the same as the processing from 1 to ST4.

In step ST 71, the word type obtaining means 71 obtains the word and the position in the word of each hypothesis currently being recognized from the likelihood calculating means 15, and the word dictionary storing means 14.
From the part of speech of the hypothesis word. Then, the duration length data changing unit 72 selects and sets the duration length data corresponding to the part of speech and the position in the word of the hypothesis acquired by the word type acquisition unit 71 from the duration length data storage unit 41. . For example, if the word currently being recognized is the particle "ga" and the phoneme position is the final / a /, there are many cases where there is a gap in the utterance, so the duration time data corresponding to the late utterance is selected. And set.

In step ST72, the likelihood calculating means 15 reads the hypothesis of each word from the word dictionary storage means 14, and stores the output probabilities of each HMM state calculated in step ST4 and the word probabilities. HMM acoustic model of the word and the duration data changing means 72
Calculates the acoustic likelihood of each hypothesis based on the set duration data and develops the hypothesis.

The processing from steps ST6 to ST10 is the same as the processing from steps ST6 to ST10 in FIG.
The processing is the same as that described above.

As described above, according to the sixth embodiment, optimal duration data is set according to the type of word such as part of speech, and duration control is performed according to the utterance situation. The effect is that the recognition accuracy can be increased even if the value changes.

Embodiment 7 FIG. FIG. 20 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 7 of the present invention. In the figure, reference numeral 41 denotes duration time data storage means for storing duration time data of utterance corresponding to the utterance speed; 81 is a language model storage means for storing word chain probabilities such as bigrams and trigrams; 82 is a language model storage means which acquires words of each hypothesis currently being recognized and words recognized before that from the likelihood calculating means 15; The word chain probability obtaining means for obtaining the chain probability of those words from 81, and the duration data corresponding to the word chain probability obtained by the word chain probability obtaining means 82 from the duration data storage means 41. This is a duration time data changing unit that is selected and set.

In this embodiment, a duration data storage unit 41, a language model storage unit 81, a word chain probability acquisition unit 82, and a duration data change unit 83 are added to the conventional configuration shown in FIG. It is.

Next, the operation will be described. First, the duration time data storage means 41 is an external storage device (not shown).
, And obtains and stores respective duration data for different utterance speeds such as fast utterance, normal utterance, and slow utterance. The language model storage means 81 acquires and stores word chain probabilities such as bigrams and trigrams from an external storage device.

FIG. 21 is a flowchart showing a process performed by the speech recognition apparatus according to the seventh embodiment of the present invention. The processing from step ST1 to ST4 is the same as the processing from step ST1 to ST4 in FIG. 27 of the related art.

In step ST 81, the word chain probability obtaining means 82 obtains the words of each hypothesis currently being recognized and the words recognized before that from the likelihood calculating means 15, and reads those words from the language model storing means 81. Get the chain probability of. Then, the duration data changing unit 83 stores the duration data corresponding to the word chain probability acquired by the word chain probability acquiring unit 82 in the duration data storage unit 41.
And set the duration time data to be used.

If the word chain probability of the word obtained by the word chain probability obtaining means 82 is close to 1, there is a small possibility that there will be a gap in the utterance. Select and set duration time data corresponding to fast utterance or normal utterance. On the other hand, when the word concatenation probability of the word acquired by the word concatenation probability acquisition means 82 is close to 0, there is a high possibility that only words that are not very related are continuous, and there is a possibility that a gap will occur in the utterance. Is large, the duration data changing means 83 selects and sets duration data corresponding to a normal utterance or a slow utterance.

In step ST82, the likelihood calculation means 15 reads out the hypothesis of each word from the word dictionary storage means 14, and stores the output probabilities of each HMM state calculated in step ST4 and the word dictionary storage means 14. HMM acoustic model of the word and the duration data changing means 83
Calculates the acoustic likelihood of each hypothesis based on the set duration data and develops the hypothesis.

The processing from steps ST6 to ST10 is the same as the processing from steps ST6 to ST10 in FIG.
The processing is the same as that described above.

As described above, according to the seventh embodiment, the optimal duration data is set according to the ease of word connection, and the duration control is performed according to the content of the utterance. However, even if the content of the information changes, the effect that the recognition accuracy can be improved can be obtained.

Embodiment 8 FIG. FIG. 22 is a diagram showing a configuration of a speech recognition apparatus according to Embodiment 8 of the present invention. In the figure, reference numeral 91 denotes the maximum likelihood of the remaining hypothesis from the pruning means 16, and This is a maximum likelihood hypothesis adding means for adding a hypothesis that has the first utterance. In this embodiment, a maximum likelihood hypothesis adding means 91 is added to the conventional configuration shown in FIG.

Next, the operation will be described. FIG. 23 is a flowchart showing processing of the voice recognition device according to the eighth embodiment of the present invention. The processing from step ST1 to ST6 is the same as the processing from step ST1 to ST6 in the conventional FIG.

In step ST91, the maximum likelihood hypothesis adding means 91 obtains the maximum likelihood of the remaining hypothesis from the pruning means 16 in the first N frames (N is an integer) of the utterance, and calculates the obtained maximum likelihood. Add a hypothesis that you have the first utterance.

The processing from step ST7 to ST10 is the same as the conventional processing from step ST7 to ST10 shown in FIG.

Step ST91 will be described in detail. FIG. 24 is a diagram illustrating a hypothesis developed in the related art when “tomorrow” is uttered but expiration is applied to the first two frames. At time t = 0, the hypothesis that the phoneme position is the first / a / in the recognition candidate “Tomorrow” and the hypothesis that the phoneme position is the first / h / in the recognition candidate “Bridge” are drawn. Also, there is a hypothesis that the phoneme position is the first for each recognition candidate. As the time progresses, these hypotheses are developed as in the conventional example, and those with low likelihoods are pruned. In the figure, the thick framed hypothesis is the one that gives the maximum likelihood in that frame. In this figure, the first exhalation is recognized as "ha", and the word "hashi (bridge)" has the highest likelihood as a word, and the correct "tomorrow" is pruned on the way. .

FIG. 25 shows an example of adding a hypothesis and developing the hypothesis in this embodiment. At time t = 0, in addition to the phoneme position / h / at the recognition candidate "bridge" shown in the figure, a hypothesis whose phoneme position is the first in each recognition candidate exists. At time t = 1, each hypothesis is developed.
, It is assumed that the likelihood of the bold frame hypothesis is the maximum, and the likelihood at that time is A. At this time, the maximum likelihood hypothesis adding means 91 adds a hypothesis whose phoneme position is first and the likelihood is A. In the figure, at t = 1, the hypothesis that the recognition candidate “tomorrow” is the phoneme position / a / and the likelihood is A, and the hypothesis that the recognition candidate “school” is the phoneme position / g / and the likelihood is A is Has been added. Other hypotheses to be added are added for each recognition candidate such as the hypothesis of the phoneme position / k / likelihood A in the recognition candidate “article”.

Similarly, the maximum likelihood at time t = 2 is B
, The phoneme position / a /,
The hypothesis of likelihood B, the phoneme position / g /, the hypothesis of likelihood B, etc. are added for the recognition candidate “school”.

Recognition candidate "tomorrow" added at time t = 2
Is not affected by the first exhalation and has a large likelihood, so it is not affected by pruning and is not rejected. As a result, the correct hypothesis is developed and the recognition accuracy is improved. .

Of the added hypotheses, those of recognition candidates that are not the correct answer (such as (A2) and (B2)) will not match the acoustic model as the time advances, and the likelihood will relatively decrease. It does not affect recognition. Similarly, the recognition candidate “bridge” having the maximum likelihood in the expiration part also becomes rejected because the acoustic model does not match as the time elapses, and the likelihood decreases relatively.

As described above, according to the eighth embodiment, the first N frames of the utterance always have the maximum likelihood,
Since we add a hypothesis with the first phoneme position of the recognized word,
Even if there is a hypothesis that recognizes the first exhalation of a utterance as a word, the effect that the recognition accuracy can be improved without being affected by the exhalation or the like is obtained.

The speech recognition processing in each of the above embodiments is realized by a speech recognition program, and the speech recognition program is provided by being recorded on a recording medium.

[0122]

As described above, according to the present invention, a simple acoustic model storage means for storing a simple acoustic model of each phoneme, an acoustic feature vector output from the acoustic analysis means, and a simple acoustic model storage means A simple acoustic model probability calculating means for calculating a simple acoustic output probability of each HMM state at each time within a predetermined time including the current time by using a simple acoustic model of each phoneme stored in The means determines the order of the simplified sound output probability of each HMM state at each time, calculates the order variation width of each HMM state within a predetermined time including the current time,
A rank variation calculator for calculating an average of the rank variation of the state, and adjusting a parameter related to speech recognition based on the average of the rank variation calculated by the rank variation calculator, thereby obtaining a simple sound of each phoneme. Since the parameters relating to speech recognition are adjusted using the model, there is an effect that the recognition accuracy can be increased without increasing the calculation cost so much.

According to the present invention, there is provided a beam width changing means for setting a beam width as a parameter relating to speech recognition, and when the average of the order variation width calculated by the order variation calculating means is smaller than a predetermined value, the beam width is changed. The changing means sets the beam width small, and when the average of the order variation width is larger than a predetermined value, the beam width changing means sets the beam width large, and the pruning means sets the beam width to the beam width set by the beam width changing means. By rejecting the hypothesis, the beam width is set appropriately using a simple acoustic model of each phoneme.
There is an effect that the calculation cost does not increase so much, the correct answer is suppressed from being pruned, and the recognition accuracy can be increased.

According to the present invention, there is provided the language weight changing means for setting the weight of the language likelihood added to the likelihood as a parameter relating to speech recognition, and the average of the order variation width calculated by the order variation calculating means is provided. Is smaller than a predetermined value, the language weight changing means sets a small weight of the language likelihood, and if the average of the rank variation is larger than a predetermined value, the language weight changing means sets a large weight of the language likelihood. Then, the likelihood calculating means calculates the likelihood of the hypothesis based on the weight of the language likelihood set by the language weight changing means, and appropriately sets the language weight using a simple acoustic model of each phoneme. So
There is an effect that the recognition accuracy can be increased without significantly increasing the calculation cost.

According to the present invention, duration data storage means for storing duration data of utterances corresponding to a plurality of utterance speeds, and duration setting for setting duration data as a parameter relating to speech recognition. Data change means, and when the average of the order change width calculated by the order change calculation means is smaller than a predetermined value, the duration time data change means utters the voice data lately stored in the duration time data storage means. When the average of the order variation width is larger than a predetermined value, the duration data changing unit outputs the uttered duration data stored in the duration data storage unit. The likelihood calculation means,
Based on the duration data selected by the duration data changing unit, the likelihood of the hypothesis is calculated, and the duration data corresponding to the utterance speed is set using a simple acoustic model of each phoneme. Therefore, there is an effect that the recognition accuracy can be increased even if the speed of utterance changes without increasing the calculation cost so much.

According to the present invention, there is provided thinning-out determining means for determining whether or not to perform frame processing as a recognition processing at each time, as a parameter relating to speech recognition. When the average is smaller than a predetermined value, the thinning-out determining means sets the thinning out of the frame processing to be present, and when the average of the order variation width is larger than the predetermined value, the thinning-out determining means sets no thinning out of the frame processing. As a result, the use of a simple acoustic model of each phoneme determines whether or not to perform thinning in accordance with the situation of voice data. Therefore, there is an effect that the calculation accuracy does not increase so much and the recognition accuracy can be increased.

According to the present invention, there is provided a thinning-out rate determining means for determining a thinning-out rate of a frame process which is a recognition process at each time, as a parameter relating to voice recognition, and a rank variation width calculated by the rank variation calculating means is provided. When the average is smaller than the predetermined value, the thinning rate determining means sets the thinning rate of the frame processing to be large, and when the average of the order variation width is larger than the predetermined value, the thinning rate determining means sets the thinning rate of the frame processing to be small. Thus, since the thinning rate according to the situation of the audio data is determined using the simple acoustic model, there is an effect that the recognition accuracy can be increased without increasing the calculation cost so much.

According to the present invention, the duration data storage means for storing the duration data of utterances corresponding to a plurality of utterance speeds, and the word and the intra-word position of each hypothesis currently being recognized by the likelihood calculating means. Word type acquiring means for acquiring the part of speech of the word of the hypothesis from the word dictionary storage means, and the duration data corresponding to the part of speech and the position in the word of the hypothetical word acquired by the word type acquiring means. Length data changing means selected from the long data storage means,
The likelihood calculating means calculates the likelihood of the hypothesis based on the duration data selected by the duration data changing means, thereby setting the optimal duration data according to the type of word such as part of speech, and uttering. Since the duration control is performed according to the situation described above, there is an effect that the recognition accuracy can be increased even if the utterance speed changes.

According to the present invention, the language model storage means for storing the word chain probability, the duration data storage means for storing the duration data of the utterance corresponding to the utterance speed, and A word chain probability obtaining unit that obtains a word of each hypothesis being recognized and a word recognized before that, and obtains a chain probability of those words from a language model storage unit; Comprising a duration data changing means for selecting duration data according to the chain probability from the duration data storage means,
Based on the duration data selected by the duration data changing means, the likelihood calculating means calculates the likelihood of the hypothesis, thereby setting the optimal duration data according to the ease of word connection, Since the duration control is performed according to the content of the utterance, there is an effect that the recognition accuracy can be increased even if the content of the utterance changes.

According to the present invention, the maximum likelihood of the hypothesis is acquired from the pruning means, and the maximum likelihood hypothesis having the acquired maximum likelihood and adding the hypothesis as the first utterance and outputting to the pruning means is added. With the provision of the means, a hypothesis having the maximum likelihood and having the first phoneme position of the recognized word is added, so even if there is a hypothesis that recognizes the first exhalation of speech as a word, the There is an effect that recognition accuracy can be increased without being affected.

According to the present invention, the first step of taking in audio data at predetermined time points and performing acoustic analysis to output an acoustic feature vector, and the acoustic feature vector output in the first step, are stored in advance. A second step of calculating a simple acoustic output probability of each HMM state at each time within a predetermined time including a current time by using a simple acoustic model of each phoneme, and a simple acoustic output probability of each HMM state at each time A third step of calculating the order fluctuation width of each HMM state within a predetermined time including the current time, and calculating the average of the order fluctuation width of the HMM state; and The fourth step of setting the beam width of the sound, the acoustic feature vector output in the first step, the acoustic model of each phoneme and the acoustic model of a word stored in advance. A fifth step of calculating the likelihood of the hypothesis that is a recognition candidate, and obtaining the maximum likelihood from the calculated likelihood of the hypothesis, and using the obtained maximum likelihood to be equal to or less than the predetermined beam width set in the fourth step. By recognizing the speech by providing a sixth step of rejecting the hypothesis of the above and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate, a simple acoustic model of each phoneme is used. Since the beam width is set appropriately, there is an effect that the calculation cost does not increase so much, the correct answer can be suppressed from being pruned, and the recognition accuracy can be increased.

According to the present invention, the first step of taking in audio data at predetermined time points and performing acoustic analysis to output an acoustic feature vector, and the acoustic feature vector output in the first step, are stored in advance. A second step of calculating a simple acoustic output probability of each HMM state at each time within a predetermined time including a current time by using a simple acoustic model of each phoneme, and a simple acoustic output probability of each HMM state at each time A third step of calculating the order variation of each HMM state within a predetermined time including the current time, calculating an average of the order variation of all the HMM states, and based on the average of the order variation. A fourth step of adjusting parameters related to speech recognition, an acoustic feature vector output in the first step, an acoustic model of each phoneme and a sound of a word stored in advance. A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the model and parameters related to the speech recognition adjusted in the fourth step; and obtaining the maximum likelihood from the calculated likelihood of the hypothesis. A sixth step of rejecting a hypothesis having a predetermined beam width or less from the likelihood, and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate.
By using the simple acoustic model of each phoneme to adjust the parameters related to speech recognition by recognizing the speech with the steps of, it is possible to increase the recognition accuracy without significantly increasing the calculation cost. There is an effect that can be.

According to the present invention, the first step of taking in audio data at predetermined time points and performing acoustic analysis to output an acoustic feature vector, and the acoustic feature vector output in the first step, are stored in advance. A second step of calculating the output probability of each HMM state from the acoustic model of each phoneme that is present, acquiring the word and position in the word of each hypothesis currently being recognized, and acquiring the part of speech of the previously stored hypothesis word The third step to perform, and the duration of time data corresponding to the part of speech of the word of the hypothesis acquired in the third step and the position within the word,
A fourth step of selecting from utterance duration data corresponding to the utterance speed stored in advance, an output probability of each HMM state calculated in the second step, and a sound of a word stored in advance. A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the model and the duration data selected in the fourth step; and obtaining the maximum likelihood from the calculated likelihood of the hypothesis. By recognizing speech by including a sixth step of rejecting a hypothesis having a beam width equal to or smaller than a predetermined value and a seventh step of outputting the hypothesis remaining in the sixth step as a recognition candidate, Since the optimal duration data is set according to the type of the word and the duration control is performed according to the utterance situation, the recognition accuracy can be improved even if the utterance speed changes. is there.

According to the present invention, the first step of taking in audio data at predetermined time points and performing acoustic analysis to output an acoustic feature vector, and the acoustic feature vector output in the first step are stored in advance. A second step of calculating an output probability of each HMM state from an acoustic model of each phoneme; acquiring words of each hypothesis currently being recognized and words recognized before that; ,
A third step of obtaining the chain probabilities of those words;
A fourth step of selecting duration time data corresponding to the word chain probability obtained in the third step from utterance duration data corresponding to the utterance speed stored in advance; and a second step. A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the calculated output probabilities of each HMM state, the acoustic model of the word stored in advance, and the duration data selected in the fourth step; A sixth step of obtaining a maximum likelihood from the calculated likelihood of the hypothesis and rejecting a hypothesis having a predetermined beam width or less from the obtained maximum likelihood;
And a seventh step of outputting a hypothesis left in the sixth step as a recognition candidate. By recognizing the speech, optimal duration time data is set according to the ease of connecting words, and Since the duration control is performed according to the content, the recognition accuracy can be improved even if the content of the utterance changes.

According to the present invention, the first step of taking in audio data at predetermined time points and performing acoustic analysis to output an acoustic feature vector, and the acoustic feature vector output in the first step, are stored in advance. A second step of calculating the likelihood of a hypothesis that is a recognition candidate based on the acoustic model of each phoneme and the acoustic model of the word, and a third step of calculating the maximum likelihood from the likelihood of the hypothesis calculated in the second step. A fourth step of obtaining the maximum likelihood obtained in the step and the third step, and adding a hypothesis having the obtained maximum likelihood and being the first utterance
And a fifth step of rejecting a hypothesis of a predetermined beam width or less from the maximum likelihood obtained in the third step;
And a sixth step of outputting the hypothesis left in the fifth step as a recognition candidate. By recognizing the speech, a hypothesis having the maximum likelihood and having the first phoneme position of the recognized word is added. Even if there is a hypothesis that recognizes the first exhalation of speech as a word, there is an effect that recognition accuracy can be increased without being affected by the exhalation or the like.

[Brief description of the drawings]

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

FIG. 2 is a flowchart showing processing of the voice recognition device according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing a process for setting a beam width according to the first embodiment of the present invention.

FIG. 4 is a diagram showing a simplified sound output probability at each time in the HMM state according to the first embodiment of the present invention.

FIG. 5 is a diagram showing a simplified sound output probability at each time in the HMM state according to the first embodiment of the present invention.

FIG. 6 is a diagram showing a configuration of a voice recognition device according to a second embodiment of the present invention.

FIG. 7 is a flowchart showing processing of the voice recognition device according to the second embodiment of the present invention.

FIG. 8 is a flowchart showing a process for setting a weight of language likelihood according to the second embodiment of the present invention.

FIG. 9 is a diagram showing a configuration of a voice recognition device according to a third embodiment of the present invention.

FIG. 10 is a flowchart showing processing of the voice recognition device according to the third embodiment of the present invention.

FIG. 11 is a flowchart showing a process for setting duration time data according to the third embodiment of the present invention.

FIG. 12 is a diagram showing a configuration of a voice recognition device according to a fourth embodiment of the present invention.

FIG. 13 is a flowchart showing processing of the voice recognition device according to the fourth embodiment of the present invention.

FIG. 14 is a flowchart showing a process for determining whether or not to omit the frame process according to the fourth embodiment of the present invention.

FIG. 15 is a diagram showing a configuration of a voice recognition device according to a fifth embodiment of the present invention.

FIG. 16 is a flowchart showing processing of the speech recognition device according to the fifth embodiment of the present invention.

FIG. 17 is a flowchart showing a process for determining a thinning rate of frame processing according to the fifth embodiment of the present invention.

FIG. 18 is a diagram showing a configuration of a voice recognition device according to a sixth embodiment of the present invention.

FIG. 19 is a flowchart showing processing of the voice recognition device according to the sixth embodiment of the present invention.

FIG. 20 is a diagram showing a configuration of a speech recognition device according to a seventh embodiment of the present invention.

FIG. 21 is a flowchart showing processing of the voice recognition device according to the seventh embodiment of the present invention.

FIG. 22 is a diagram showing a configuration of a speech recognition device according to an eighth embodiment of the present invention.

FIG. 23 is a flowchart showing processing of the voice recognition device according to the eighth embodiment of the present invention.

FIG. 24 is a diagram showing a conventional hypothesis developed when exhalation is applied.

FIG. 25 is a diagram showing a hypothesis developed when exhalation is applied according to the eighth embodiment of the present invention.

FIG. 26 is a diagram showing a configuration of a conventional voice recognition device.

FIG. 27 is a flowchart showing processing of a conventional speech recognition device.

FIG. 28 is a diagram showing a specific example of a conventional word dictionary.

FIG. 29 is a diagram showing an example of the structure of a conventional word acoustic model.

FIG. 30 is a diagram illustrating a conventional situation in which a hypothesis is developed as time advances.

FIG. 31 is a diagram illustrating an example of a conventional duration time.

[Explanation of symbols]

11 sound data storage means, 12 sound analysis means, 13
Acoustic model storage means, 14 Word dictionary storage means, 15
Likelihood calculating means, 16 pruning means, 17 recognition result output means, 21 simple acoustic model storing means, 22 simple acoustic model probability calculating means, 23 rank variation calculating means, 24
Beam width changing means, 31 Language weight changing means, 41
Duration time data storage means, 42 duration data change means, 51 decimation determination means, 61 decimation rate determination means, 71 word type acquisition means, 72 duration data change means, 81 language model storage means, 82 word chain probability Acquisition means, 83 Duration length data change means, 91
Means of maximum likelihood hypothesis addition.

Claims (19)

[Claims] 1. An audio data storage means for converting input audio into digital data and storing the digital data as audio data, an audio analysis means for taking in the audio data at predetermined time intervals, performing an acoustic analysis and outputting an acoustic feature vector, Acoustic model storage means for storing an acoustic model of, a word dictionary storage means for generating and storing an acoustic model of a word from a phoneme description of each word in a word dictionary, and an acoustic feature vector output from the acoustic analysis means,
An acoustic model of each phoneme stored in the acoustic model storage means, and a likelihood calculation means for calculating the likelihood of a hypothesis that is a recognition candidate, based on the acoustic model of the word stored in the word dictionary storage means, Pruning means for obtaining the maximum likelihood from the likelihood of the hypothesis calculated by the likelihood calculating means, and rejecting a hypothesis having a predetermined beam width or less from the obtained maximum likelihood, and a hypothesis left by the pruning means. In a speech recognition device including a recognition result output unit that outputs a recognition candidate, a simple acoustic model storage unit that stores a simple acoustic model of each phoneme; an acoustic feature vector output from the acoustic analysis unit;
A simple acoustic model probability calculating means for calculating a simple acoustic output probability of each HMM state at each time within a predetermined time including the current time by a simple acoustic model of each phoneme stored in the simple acoustic model storage means; Each H at each time obtained by the simple acoustic model probability calculation means.
Rank fluctuation calculating means for calculating the rank of the simplified sound output probability of the MM state, calculating the rank fluctuation width of each HMM state within a predetermined time including the current time, and calculating the average of the rank fluctuation width of the HMM state. A speech recognition device for adjusting a parameter relating to speech recognition based on an average of the order variation width calculated by the order variation calculating means. 2. A beam width changing means for setting a beam width as a parameter relating to speech recognition, wherein the beam width changing means calculates a beam width when the average of the rank fluctuation width calculated by the rank fluctuation calculating means is smaller than a predetermined value. The beam width is set small, and when the average of the order variation width is larger than a predetermined value, the beam width changing means sets the beam width large, and the pruning means sets the beam set by the beam width changing means. 2. The hypothesis rejected based on the width.
The speech recognition device according to the above. 3. A language weight changing means for setting a weight of a language likelihood added to the likelihood as a parameter relating to speech recognition, wherein an average of a rank variation width calculated by the rank variation calculating means is larger than a predetermined value. When the language weight is small, the language weight changing means sets the weight of the language likelihood small, and when the average of the order variation width is larger than a predetermined value, the language weight changing means sets the weight of the language likelihood large. 2. The speech recognition apparatus according to claim 1, wherein the likelihood calculating means calculates the likelihood of the hypothesis based on the weight of the language likelihood set by the language weight changing means. 4. A duration data storage unit for storing duration data of utterance corresponding to a plurality of utterance speeds, and a duration data changing unit for setting duration data as a parameter relating to speech recognition. When the average of the order change width calculated by the order change calculating means is smaller than a predetermined value, the duration time data changing means is:
The duration time data uttered late stored in the duration time data storage means is selected, and when the average of the rank variation width is larger than a predetermined value, the duration time data changing means selects the duration time data. The duration utterance data that has been uttered earlier stored in the duration data storage is selected, and the likelihood calculation means determines the likelihood of the hypothesis based on the duration data selected by the duration data changer. The speech recognition device according to claim 1, wherein the speech recognition device performs a calculation. 5. A thinning-out determining means for determining whether or not a frame processing which is a recognition processing at each time is thinned out as a parameter relating to voice recognition, wherein an average of the rank variation calculated by the rank variation calculating means is a predetermined value. If it is smaller, the thinning-out determining means sets the thinning-out of the frame processing to "yes", and if the average of the order variation width is larger than a predetermined value, the thinning-out determining means sets the thinning-out of the frame processing to no. Claim 1.
The speech recognition device according to the above. 6. A thinning-out rate determining means for determining a thinning-out rate of frame processing as a recognition processing at each time as a parameter relating to voice recognition, wherein the average of the rank variation calculated by the rank variation calculating means is a predetermined value. If it is smaller, the thinning rate determining means sets the thinning rate of the frame processing to be large, and if the average of the order variation width is larger than a predetermined value, the thinning rate determining means sets the thinning rate of the frame processing to be small. The speech recognition device according to claim 1, wherein: 7. Audio data storage means for converting input speech into digital data and storing the same as speech data; acoustic analysis means for taking in the speech data at predetermined time points and performing acoustic analysis to output an acoustic feature vector; Acoustic model storage means for storing an acoustic model of, a word dictionary storage means for generating and storing an acoustic model of a word from a phoneme description of each word in a word dictionary, and an acoustic feature vector output from the acoustic analysis means,
An acoustic model of each phoneme stored in the acoustic model storage means, and a likelihood calculation means for calculating the likelihood of a hypothesis that is a recognition candidate, based on the acoustic model of the word stored in the word dictionary storage means, Pruning means for obtaining the maximum likelihood from the likelihood of the hypothesis calculated by the above-mentioned likelihood calculating means, and rejecting a hypothesis having a predetermined beam width or less from the obtained maximum likelihood, and a hypothesis left by the pruning means. A speech recognition device comprising: a recognition result output unit that outputs a recognition candidate; a duration data storage unit that stores duration data of utterances corresponding to a plurality of utterance speeds; A word type obtaining unit that obtains the word of each hypothesis and the position within the word, and obtains the part of speech of the word of the hypothesis from the word dictionary storage unit; and a word product of the hypothesis obtained by the word type obtaining unit. And duration time data changing means for selecting duration time data corresponding to the position within the word from the duration data storage means, wherein the likelihood calculating means selects the duration data changing means. A speech recognition apparatus for calculating a likelihood of a hypothesis based on duration time data. 8. Audio data storage means for converting input speech into digital data and storing the same as speech data, acoustic analysis means for taking in the speech data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, and each phoneme. Acoustic model storage means for storing an acoustic model of, a word dictionary storage means for generating and storing an acoustic model of a word from a phoneme description of each word in a word dictionary, and an acoustic feature vector output from the acoustic analysis means,
An acoustic model of each phoneme stored in the acoustic model storage means, and a likelihood calculation means for calculating the likelihood of a hypothesis that is a recognition candidate, based on the acoustic model of the word stored in the word dictionary storage means, Pruning means for obtaining the maximum likelihood from the likelihood of the hypothesis calculated by the above-mentioned likelihood calculating means, and rejecting a hypothesis having a predetermined beam width or less from the obtained maximum likelihood, and a hypothesis left by the pruning means. A speech recognition apparatus comprising: a recognition result output unit that outputs a recognition candidate; a language model storage unit that stores a word chain probability; and a duration data that stores duration data of an utterance corresponding to an utterance speed. Means, a word sequence for obtaining words of each hypothesis currently being recognized and words recognized before that from the likelihood calculating means, and obtaining a chain probability of those words from the language model storage means. A probability obtaining means, and a duration data changing means for selecting duration data according to the word chain probability obtained by the word chain probability obtaining means from the duration data storage means; A speech recognition device, wherein a calculating means calculates a likelihood of a hypothesis based on the duration data selected by the duration data changing means. 9. Voice data storage means for converting input voice into digital data and storing the voice data as voice data, acoustic analysis means for capturing the voice data at predetermined time points and performing acoustic analysis to output an acoustic feature vector; Acoustic model storage means for storing an acoustic model of, a word dictionary storage means for generating and storing an acoustic model of a word from a phoneme description of each word in a word dictionary, and an acoustic feature vector output from the acoustic analysis means,
An acoustic model of each phoneme stored in the acoustic model storage means, and a likelihood calculation means for calculating the likelihood of a hypothesis that is a recognition candidate, based on the acoustic model of the word stored in the word dictionary storage means, Pruning means for obtaining the maximum likelihood from the likelihood of the hypothesis calculated by the above-mentioned likelihood calculating means, and rejecting a hypothesis having a predetermined beam width or less from the obtained maximum likelihood, and a hypothesis left by the pruning means. A speech recognition apparatus comprising: a recognition result output unit that outputs a recognition candidate; and obtains a maximum likelihood of a hypothesis from the pruning unit, adds the obtained maximum likelihood, and adds a hypothesis that is the first utterance. A speech recognition device comprising a maximum likelihood hypothesis adding unit that outputs to a pruning unit. 10. A first step of taking in audio data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, an acoustic feature vector output in the first step,
By a simple acoustic model of each phoneme stored in advance,
Each HM at each time within a predetermined time including the current time
A second step of calculating a simple sound output probability of the M state; obtaining a rank of the simple sound output probability of each HMM state at each of the above times;
A third step of calculating an order variation width of the state and calculating an average of the order variation width of the HMM state; a fourth step of setting a predetermined beam width based on the average of the order variation width; The acoustic feature vector output in the step of
Fifth calculation of the likelihood of a hypothesis that is a recognition candidate is performed based on the acoustic model of each phoneme and the acoustic model of a word stored in advance.
And a sixth step of obtaining the maximum likelihood from the calculated likelihood of the hypothesis, and rejecting a hypothesis having a beam width equal to or smaller than the predetermined beam width set in the fourth step from the obtained maximum likelihood; And a seventh step of outputting the hypothesis left in step 6 as a recognition candidate. 11. A first step of taking in audio data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, an acoustic feature vector output in the first step,
By a simple acoustic model of each phoneme stored in advance,
Each HM at each time within a predetermined time including the current time
A second step of calculating a simple sound output probability of the M state; obtaining a rank of the simple sound output probability of each HMM state at each of the above times;
A third step of calculating the order change width of the states and calculating an average of the order change ranges of all the HMM states; a fourth step of adjusting a parameter relating to speech recognition based on the average of the order change ranges; The acoustic feature vector output in the first step;
A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the acoustic model of each phoneme and the acoustic model of the word stored in advance and the parameters related to the speech recognition adjusted in the fourth step; A sixth step of obtaining a maximum likelihood from the likelihood of the obtained hypothesis, and rejecting a hypothesis of a predetermined beam width or less from the obtained maximum likelihood; and outputting the hypothesis left in the sixth step as a recognition candidate. And a seventh step of recognizing a voice. 12. A first step of taking in audio data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, an acoustic feature vector output in the first step,
A second step of calculating an output probability of each HMM state from a pre-stored acoustic model of each phoneme; acquiring a word of each hypothesis currently recognized and a position in the word; A third step of acquiring the part of speech of the hypothesis, and the duration time data corresponding to the part of speech and the position in the word of the hypothesis word acquired in the third step, by continuing the utterance corresponding to the utterance speed stored in advance. A fourth step of selecting from time length data; an output probability of each HMM state calculated in the second step; an acoustic model of a word stored in advance;
A fifth step of calculating the likelihood of the hypothesis that is a recognition candidate based on the duration data selected in the step; and obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and a predetermined beam from the obtained maximum likelihood. A speech recognition method comprising: recognizing speech, comprising: a sixth step of rejecting a hypothesis having a width equal to or less than a width; and a seventh step of outputting a hypothesis left in the sixth step as a recognition candidate. 13. A first step of taking in audio data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, and the acoustic feature vector output in the first step and a prestored phoneme of each phoneme. A second step of calculating the output probabilities of each HMM state from the acoustic model, acquiring the words of each hypothesis currently being recognized and the words recognized before, and calculating them from the chain probability of the words stored in advance. A third step of acquiring the chain probability of the word, and the duration data corresponding to the chain probability of the word acquired in the third step are converted into the duration data of the utterance corresponding to the utterance speed stored in advance. A fourth step of selecting from the following: an output probability of each HMM state calculated in the second step; an acoustic model of a word stored in advance;
A fifth step of calculating the likelihood of the hypothesis that is a recognition candidate based on the duration data selected in the step; and obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and a predetermined beam from the obtained maximum likelihood. A speech recognition method comprising: recognizing speech, comprising: a sixth step of rejecting a hypothesis having a width equal to or less than a width; and a seventh step of outputting a hypothesis left in the sixth step as a recognition candidate. 14. A first step of taking in audio data at predetermined time points, performing acoustic analysis and outputting an acoustic feature vector, and an acoustic feature vector output in the first step;
A second calculating unit that calculates a likelihood of a hypothesis that is a recognition candidate using an acoustic model of each phoneme and an acoustic model of a word stored in advance.
And a third step of obtaining the maximum likelihood from the likelihood of the hypothesis calculated in the second step; and obtaining the maximum likelihood obtained in the third step. Add a hypothesis to be the first to utter
A fifth step of rejecting a hypothesis having a predetermined beam width or less from the maximum likelihood obtained in the third step, and a sixth step of outputting the hypothesis left in the fifth step as a recognition candidate. A voice recognition method comprising the steps of: 15. A first step of taking in audio data at predetermined times and performing acoustic analysis to output an acoustic feature vector; and an acoustic feature vector output in the first step.
By a simple acoustic model of each phoneme stored in advance,
Each HM at each time within a predetermined time including the current time
A second step of calculating a simple sound output probability of the M state; obtaining a rank of the simple sound output probability of each HMM state at each of the above times;
A third step of calculating an order variation width of the state and calculating an average of the order variation width of the HMM state; a fourth step of setting a predetermined beam width based on the average of the order variation width; The acoustic feature vector output in the step of
Fifth calculation of the likelihood of a hypothesis that is a recognition candidate is performed based on the acoustic model of each phoneme and the acoustic model of a word stored in advance.
And a sixth step of obtaining the maximum likelihood from the calculated likelihood of the hypothesis, and rejecting a hypothesis having a beam width equal to or smaller than the predetermined beam width set in the fourth step from the obtained maximum likelihood; A recording medium on which a speech recognition program for causing a computer to execute a seventh step of outputting a hypothesis left in step 6 as a recognition candidate is recorded. 16. A first step of taking in audio data at predetermined times and performing acoustic analysis to output an acoustic feature vector, an acoustic feature vector output in the first step,
By a simple acoustic model of each phoneme stored in advance,
Each HM at each time within a predetermined time including the current time
A second step of calculating a simple sound output probability of the M state; obtaining a rank of the simple sound output probability of each HMM state at each of the above times;
A third step of calculating the order change width of the states and calculating an average of the order change ranges of all the HMM states; a fourth step of adjusting a parameter relating to speech recognition based on the average of the order change ranges; The acoustic feature vector output in the first step;
A fifth step of calculating the likelihood of a hypothesis that is a recognition candidate based on the acoustic model of each phoneme and the acoustic model of the word stored in advance and the parameters related to the speech recognition adjusted in the fourth step; A sixth step of obtaining a maximum likelihood from the likelihood of the obtained hypothesis, and rejecting a hypothesis of a predetermined beam width or less from the obtained maximum likelihood; and outputting the hypothesis left in the sixth step as a recognition candidate. A recording medium recording a voice recognition program for causing a computer to execute the seventh step. 17. A first step of taking in audio data at predetermined times, performing acoustic analysis and outputting an acoustic feature vector, an acoustic feature vector output in the first step,
A second step of calculating an output probability of each HMM state from a pre-stored acoustic model of each phoneme; acquiring a word of each hypothesis currently recognized and a position in the word; A third step of acquiring the part of speech of the hypothesis, and the duration time data corresponding to the part of speech and the position in the word of the hypothesis word acquired in the third step, by continuing the utterance corresponding to the utterance speed stored in advance. A fourth step of selecting from time length data, an output probability of each HMM state calculated in the second step, an acoustic model of a word stored in advance,
A fifth step of calculating the likelihood of the hypothesis that is a recognition candidate based on the duration data selected in the step; and obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and a predetermined beam from the obtained maximum likelihood. A recording medium storing a speech recognition program for causing a computer to execute a sixth step of rejecting a hypothesis having a width equal to or less than a width and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate. 18. A first step of taking in audio data at predetermined time points and performing acoustic analysis to output an acoustic feature vector, an acoustic feature vector output in the first step,
A second step of calculating an output probability of each HMM state from an acoustic model of each phoneme stored in advance; acquiring words of each hypothesis currently being recognized and words recognized before that; A third step of acquiring the chain probabilities of the words from the chain probabilities of the words; and converting the duration time data corresponding to the chain probabilities of the words acquired in the third step to a speech rate stored in advance. A fourth step of selecting from the corresponding utterance duration data, an output probability of each HMM state calculated in the second step, an acoustic model of a word stored in advance,
A fifth step of calculating the likelihood of the hypothesis that is a recognition candidate based on the duration data selected in the step; and obtaining a maximum likelihood from the calculated likelihood of the hypothesis, and a predetermined beam from the obtained maximum likelihood. A recording medium storing a speech recognition program for causing a computer to execute a sixth step of rejecting a hypothesis having a width equal to or less than a width and a seventh step of outputting the hypothesis left in the sixth step as a recognition candidate. 19. A first step of taking in audio data at predetermined times and performing acoustic analysis to output an acoustic feature vector; and an acoustic feature vector output in the first step.
A second calculating unit that calculates a likelihood of a hypothesis that is a recognition candidate using an acoustic model of each phoneme and an acoustic model of a word stored in advance.
And a third step of obtaining the maximum likelihood from the likelihood of the hypothesis calculated in the second step; and obtaining the maximum likelihood obtained in the third step. Add a hypothesis to be the first to utter
A fifth step of rejecting a hypothesis having a predetermined beam width or less from the maximum likelihood obtained in the third step, and a sixth step of outputting the hypothesis left in the fifth step as a recognition candidate. Recording medium on which a voice recognition program for causing a computer to execute the steps of the above is recorded.