JPH11194792A - Speech recognition device, speech recognition method, and recording medium recording the method - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To realize speech recognition of a large vocabulary with a small amount of processing and no deterioration in recognition accuracy. SOLUTION: A voice input from voice input means 1 is
The speech analysis unit 2 analyzes and outputs a time series of feature vectors, and the speech detection unit 3 determines a speech section. The sound collating unit 4 collates the time series of the feature vector with the standard pattern for the basic voice unit stored in the standard pattern storing unit 5, and the word evaluating unit 6 evaluates the recognition target based on the collation result. I do. The sound matching means 4 matches each standard pattern over the entire section of the feature vector time series of the input speech, and obtains a matching result as a time series for each standard pattern. The word evaluation means 6 evaluates each word based on the matching result for each standard pattern obtained as the time series and the information of the word dictionary 7 which describes the words of the recognition target word as a sequence of basic speech units. A recognition result is obtained according to the evaluation result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition technology in which basic units in speech language expression, such as syllables and phonemes (consonants, vowels), are used as standard patterns. Word / Sentence Speech Recognition Apparatus for Realizing Speech Recognition of Large Vocabulary with Small Processing Amount in Speech Recognition Consisting of Appearance Probability Distributions The present invention relates to a computer-readable recording medium on which recording is performed.

[0002]

2. Description of the Related Art In a speech recognition apparatus, in particular, a speech recognition apparatus in which a standard pattern is composed of a distribution of appearance probabilities of feature vectors, probability calculation occupies most of the recognition processing. In a normal speech recognition method, the number of times of the probability calculation is proportional to the number of words to be recognized. In the case of speech recognition with a large vocabulary, an enormous amount of processing is required, and a large amount of processing is required to realize real-time speech recognition. Hardware configuration was required. Several methods have been conventionally proposed as methods for reducing the enormous amount of processing required for such speech recognition. Hereinafter, a typical processing amount reduction method will be described.

As a first prior art, there is a technique called "beam search" (Transactions of the Institute of Electronics, Information and Communication Engineers, D Vol. J7).
1-D No.9 (September 1988) PP.1650-1659). The “beam search” is a method in which, among candidates to be recognized, a candidate determined to have a low possibility in the calculation process is stopped in the middle of the calculation. Among the recognition target candidates, there are a method of calculating only a certain number of candidates from the most likely one, and a method of setting a threshold value for the recognition calculation and calculating only candidates having a threshold value or more. In any case, the calculation is reduced at a fixed rate with respect to the calculation for the entire recognition candidate.

[0004] In contrast to the "beam search" in which the calculation is interrupted halfway, a method of calculating all the candidates to the end is a technique called "full search". In the case of “full search”, the calculation is performed for all the candidates to the end, so that it is guaranteed that an optimum solution is obtained. Even if a candidate is determined to be unlikely in the middle of a calculation, if the calculation is continued to the end, it may be the first correct answer. In the case of "beam search" where the calculation is terminated in the middle, the optimal solution is guaranteed. Not done.

[0005] As a second prior art, there is an example in which processing at an acoustic level is first performed to obtain a phoneme or syllable recognition result, and the result is subjected to language processing to obtain a final recognition result ( 16th Symposium on Applied Informatics Research Center "Current and Future of Speech Recognition", Applied Informatics Research Center, Tohoku University, '90 .5 29-30). In this example, recognition is performed in units of phonemes or syllables, and the results are obtained as multiple hypotheses in units of phonemes or syllables, such as phoneme lattices or syllable lattices. Ask for. The collation performed here is a collation at the symbol level, and the processing amount is significantly smaller than that of the collation processing at the acoustic level that requires probability calculation and the like. According to this method, the acoustic matching process requires only the number of phonemes or syllables, and can greatly reduce the amount of calculation. However, since the decision is made at the acoustic matching level, if the phoneme lattice or the syllable lattice does not include the correct answer candidate, no processing can be performed at the dictionary matching level to obtain a correct answer.

[0006]

As described above, the first prior art has an advantage that the processing amount can be reduced at a certain fixed rate, but the sound is proportional to the number of words to be recognized. There is a problem that the number of matching processes increases. Further, the second conventional technique has an advantage that the recognition result is obtained in units of phonemes or syllables, so that the processing amount of the sound matching processing can be suppressed to a certain processing amount. Alternatively, since a result is obtained for each syllable, there is a problem that a final result cannot be obtained for a hypothesis that has fallen from a candidate here. An object of the present invention is to solve the above problems, to obtain a final evaluation result for all the hypotheses to be recognized, and to have a fixed processing amount in which the acoustic matching processing amount is not proportional to the number of words to be recognized. A word speech recognition device and a sentence speech recognition device with a small processing amount and a small deterioration in recognition accuracy that can be suppressed, a microcomputer device therefor, a speech recognition method for words and sentences, and a computer reading the recognition method It is to provide a possible recording medium.

[0007]

In order to achieve the above object, word speech recognition according to the present invention uses syllables or syllable chains,
Or, a standard pattern for a basic voice unit such as a phoneme is collated over the entire section of the feature vector time series of the input voice, and a collation result is obtained as a time series for each standard pattern. Are evaluated based on a word dictionary described as a list of words and a collation result for each standard pattern obtained as the time series, and a recognition result is obtained.

Further, in the sentence speech recognition of the present invention, a grammar for describing a sentence to be recognized as a sequence of words is stored in advance, and a matching result for each standard pattern, a word dictionary, and each sentence based on the grammar are stored. Is evaluated to obtain a recognition result.

Further, the microcomputer device of the present invention is constituted by mounting means necessary for the recognition on a semiconductor chip, and the storage medium of the present invention comprises:
A CD-ROM or the like in which the procedure (step) for performing the word or sentence speech recognition is converted into a program code and recorded.

[0010]

Embodiments of the present invention will be described below in detail with reference to the drawings. The unit of the standard pattern is syllable,
Although various units such as syllable chains and phonemes can be considered, a case in which syllables are used as units will be described here for simplicity. In the following, mainly the case of word speech recognition will be described in detail, but it goes without saying that the present invention can be applied to sentence speech recognition. That is,
In addition to the standard pattern and word dictionary in syllable units similar to those in the following examples, using a grammar that describes a sentence to be recognized as a sequence of words, combining syllable unit matching, and based on the word dictionary and the grammar It is also possible to recognize sentences and conversation sentences uttered continuously.

FIG. 1 is a functional block diagram of one embodiment of the word speech recognition apparatus of the present invention. The input voice is first converted into an electric signal by the voice input unit 1. The voice converted into the electric signal is further analyzed by the voice analysis means 2, and a time series of feature vectors is output. The time series of this feature vector is input to the voice detection means 3.
The time series of the feature vector of the section determined to be a voice section by the voice detecting means 3 is input to the sound matching means 4. In the acoustic matching means 4, the time series of the input feature vector is compared with all the syllable standard patterns stored in the standard pattern storage means 5, and a time series of the matching result is obtained for each syllable standard pattern. The word evaluation means 6 evaluates each word based on the time series of the matching result and the word dictionary information stored in the word dictionary 7 for each syllable standard pattern, and obtains an evaluation value for each word. The determining means 8 obtains and outputs a final recognition result based on the evaluation value given to each word. For example, the words of the top five candidates with high evaluation values are output.

In the speech recognition apparatus of the present invention, it is possible to prepare a standard pattern as a time series of a feature vector and realize matching by pattern matching, or to prepare a standard pattern as a time series of an output probability distribution of a feature vector. However, matching can also be realized by probability calculation. Here, the latter method based on the output probability distribution of the feature vector,
That is, a hidden Markov model (HMM: Hidden Mar)
kov Models).

First, the HMM will be briefly described with reference to FIG. FIG. 2 is a diagram for explaining a hidden Markov model (HMM) of a basic recognition unit used in the present invention.
In the figure, each circle represents a state, and an arrow represents a transition between states. The symbol a (i, j) attached to the arrow represents the probability of a transition from state i to state j, and the symbol b (i, j, v) is the characteristic when a transition from state i to state j occurs. Represents the probability that the vector v will be output.

HMMs are roughly classified into two types, a discrete output probability distribution type HMM and a continuous output probability distribution type HMM, depending on the expression form of b (i, j, v). Discrete output probability distribution type HM
In M, the feature vector v is vector-quantized, the value of b (i, j, v) is obtained in advance for each quantization code and tabulated, and probability calculation is performed by referring to the table. In a continuous output stochastic HMM, a certain distribution function is assumed, and a feature vector v
The probability is obtained by calculating a function using. A Gaussian distribution is often used as the distribution function. When a Gaussian distribution is used, b (i, j, v) is obtained by Expression (1).

(Equation 1) Here, v, μ: column vector t: transpose ：: matrix (covariance matrix) | calculation of determinant of | Σ |: Σ

Equation (1) shows a case where the probability distribution is represented by a weighted sum of a plurality of Gaussian distributions. Although the probability distribution may be represented by a single Gaussian distribution, in general, in speech recognition of an unspecified speaker, a weighted sum of a plurality of Gaussian distributions is used. b (i, j, v) is the appearance probability (or probability density) corresponding to each state transition when the feature vector v is obtained. In the acoustic matching process, the transition probability a (ij) is further used. Calculate the cumulative probability of each state of the HMM. The cumulative probability calculation of each state can be efficiently calculated using a dynamic programming method, for example, a calculation method called a Viterbi algorithm. Formulas (2) to (4) show recurrence formulas for calculation by the Viterbi algorithm. Here, γ (i, t) is the probability of observing the feature vector time series V1, V2... Vt and being in the i-th state of the HMM.

The cumulative probability γ (i, t) in each state of the HMM can be obtained by the recurrence formula calculation of the above equations (2) to (4). A series of processes from the process of calculating the probability according to the formula (1) to the process of performing the cumulative probability calculation according to the recurrence formulas of (2) to (4) are processes performed by the sound matching means 4. The standard pattern storage means 5 stores the HMM as described in FIG. 2 for each basic syllable, syllable chain, or phoneme such as a phoneme.
When the feature vector v is obtained in step (3), the probability accumulation calculation is performed for all the HMMs stored in the standard pattern storage unit 5.

The above description is directed to a continuous HMM as the HMM.
Is a description of a case where a continuous HMM is represented by a mixture of a plurality of Gaussian distributions. However, in the following description, a case will be described in which a HMM of a type called a semi-continuous type is used as the HMM, and the processing amount is further reduced.

FIG. 3 is a detailed functional block diagram of one embodiment of the word speech recognition apparatus of the present invention when a semi-continuous HMM is used. The function of one embodiment of the word speech recognition apparatus of the present invention has been described with reference to FIG. 1. FIG. 3 shows that the acoustic matching means 4 and the standard pattern storage means 5 in FIG.
This is detailed in accordance with the MM. Semi-continuous HM
When M is used, the sound matching process is a three-stage process.
The first stage is the probability calculation in the probability calculation means 41, the second stage is the probability mixing in the probability mixing means 42, and the third step is the probability accumulation in the probability accumulation means 43. The probability calculation in the first stage probability calculation means 41 is a calculation for obtaining a probability according to each Gaussian distribution in the equation (1) from the feature vector v. Are calculated from the results of the probability calculation of the Gaussian distribution of b), and b (i, j, v) is obtained. This is a process of performing probability accumulation according to the equation.

The number of probability distributions actually present is Nu × S, where Nu is the number of recognition basic units, Su is the number of states of the HMM of the recognition basic unit, and M is the number of distributions in each state.
u × M. When Nu = 400, Su = 2, and M = 3, the number of existing probability distributions is 2400. When a semi-continuous HMM is not used, all of these probability distribution calculations must be performed. However, when a semi-continuous HMM is used, the processing amount is greatly reduced. Semi-continuous HM
In M, similar probability distributions among the 2,400 probability distributions are put together, and only the calculation of a representative probability distribution is required. For example, the above 2,400 probability distributions are clustered into 256 clusters, a representative distribution is created for each cluster, and only the calculation of the representative distribution is used instead of the calculation of the actual probability distribution. From the above, the probability calculation required 2400 times without using the semi-continuous HMM is 256 times.
I'm going to do it in times.

The representative distribution storage means 51 stores the representative distribution as described above. In this embodiment, a Gaussian distribution is used as the probability distribution, and the covariance matrix has only diagonal components. Representative distribution storage means 5
1 stores a mean vector and a covariance matrix (only diagonal components) of each Gaussian distribution. 4 is stored in the representative distribution storage unit 51.
As shown in the figure, for the representative distribution number 101, the corresponding mean vector 102, covariance matrix (only diagonal components) 1
03 is stored. Using this, the probability is calculated by the probability calculation means 41. By having such a representative distribution, each HMM has one of the representative distributions instead of having a unique probability distribution. In order to indicate any of the representative distributions, it is sufficient to know the number of the representative distribution. Therefore, each standard pattern is represented using the number of the representative distribution. Semi-continuous HMM storage means 5
2 is H described using such a representative distribution number.
MM is stored. Each semi-continuous HMM stored in the semi-continuous HMM storage means 52 is as shown in FIG.

In the probability calculating means 41, the representative distribution storing means 5
For each representative distribution stored in No. 1, the probability of each representative distribution is obtained using the feature vector v obtained by the voice detection means 3. To calculate the probability value, a Gaussian distribution formula (5) is used.

(Equation 5)

As shown in FIG. 6, the probability value calculated by the probability calculating means 41 is obtained by pairing the number 201 of the representative distribution and the probability value 202 for each representative distribution. The probability mixing means 42 performs probability mixing for all states of all HMMs stored in the semi-continuous HMM storage means 52 with reference to the probability calculation results shown in FIG. i, j, v). The probability accumulating means 43 receives the output probabilities b (i, j, v) in each state, executes a calculation by the Viterbi algorithm, and calculates and outputs the cumulative probabilities for all states of all HMMs. Note that the probability accumulation calculation performed here is continuous Viterbi calculation performed at a word spot or the like, and strictly speaking, the recurrence formula is also expressed by (2) to
It is different from the recurrence formula of (4). Further, the structure of the HMM used is slightly different from the HMM shown in FIG. FIG. 7 shows the structure of the HMM actually used. The HMM of FIG.
Is different from the HMM in that a state having no self-loop is added to the head. Although the HMM of FIG. 5 can perform only the fixed-start-point matching, the structure of FIG. 7 enables the free-end matching to be performed. The recurrence formula for performing the start-free matching is slightly different from the recurrence formulas (2) to (4).

Formulas (6) to (8) show recurrence formulas for performing start-free matching. Equation (6) is the same as equation (2), except that γ (i, t) is given 1 at each time, as in equation (7), and the maximum value is determined as in equation (8). The difference is that it is normalized by the collation path length L.

In the selection of the maximum value in the equation (8), by storing information on which state is selected, it is possible to obtain the starting point information of the collation path. In this way, the probability accumulating means 43 performs the processing shown in FIG.
The time series of the acoustic matching result as shown in FIG. As shown in FIG. 8, in the acoustic matching result time series, a probability cumulative value 302 is obtained at each time 301 as a score of each HMM, and start point information 303 of a matching path that provides such a probability cumulative value is also given. Can be FIG. 8 shows the collation result for a certain HMM. Similar collation results are obtained for all the HMMs stored in the semi-continuous HMM storage unit 52. Looking at the column of time t in FIG.
The MM is collated between the time 23 and the time t of the input voice, and it can be seen that a score of 0.009174 is obtained.

The word evaluation means 6 evaluates each word based on the time series of the matching result obtained for each of the HMMs and the word dictionary information stored in the word dictionary 7, and calculates an evaluation value for each word. Ask. FIG. 9 is a flowchart for explaining the processing performed by the word evaluation means 6. The flowchart of FIG. 9 shows the process of the word evaluation process for one word. This algorithm is a method that evaluates backward from the back syllable of the word to the front syllable. Assuming that the word to be evaluated is composed of N syllables, the syllable number i to be processed is set to N, the score is set to 0, and the search start time t is set to the end time T of the input voice (step 801).

Next, the maximum value of the time series of the matching result of the HMM corresponding to the last syllable within a certain range from the end of the input voice (time = T) is obtained. The maximum value is Smax, and the time at which the maximum value is given is tmax. Since the start time information is included in the collation result time-series information, the start time ts corresponding to tmax
A tart can be determined (step 802). i = i
−1, and the obtained Smax is added to the score of the word, and tstart is set as a new search start point t to prepare for the search for the immediately preceding syllable (step 80).
3). Steps 802 and 803 are repeated until the syllable number i to be processed becomes 0. When the syllable number i to be processed becomes 0 (step 804: Y), it means that the processing for the word has been completed, and the calculation is terminated.

The horizontal axis represents time, and the vertical axis represents H.
FIG. 10 shows a collation path on a drawing in which the state of the MM is taken (this is called a trellis). FIG. 10 shows an example of the word “Kokubunji”. Time T-
The maximum value of the collation value of the syllable “ji” is obtained from α to time T, and when the corresponding collation start point is t1, t1-
The maximum value of the collation value of the previous syllable "n" between α and t1 + α is determined. Similarly, when the matching start point corresponding to this is t2, the maximum value of the matching value of the previous syllable "bu" is determined between t2-α and t2 + α. Assuming that the matching start point corresponding to this is t3, the maximum value of the matching value of the previous syllable "ku" is determined between t3-α and t3 + α. Assuming that the matching start point corresponding to this is t4, the maximum value of the matching value of the previous syllable "ko" is determined between t4-α and t4 + α. The maximum values obtained above are accumulated and the word “Kokubunji” is accumulated.
Score.

In the above description, the time series of the sound matching result shown in FIG. 8 is obtained for all the HMMs at every time. However, since both the memory amount and the processing amount increase, the cumulative probability value exceeds a certain reference value. It is needless to say that both the memory amount and the processing amount can be reduced by recording only in the case, or by recording only the time when the cumulative probability value becomes the maximum value in the time direction. Also, in the word evaluation process shown in the flowchart of FIG. 9, it has been shown that score accumulation for all syllables is performed for all words, but when the value of Smax obtained in the middle syllable is less than a certain reference value. Needless to say, the processing amount can be reduced by terminating the processing in the middle.

The speech recognition target word is preliminarily selected by using a method with a small processing amount completely different from the speech recognition process of the present invention, and the number of target words is reduced. Of course, it is also possible to carry out the processing of.

FIG. 11 is a block diagram showing a specific hardware configuration of the speech recognition apparatus of FIG. 3 as an example of the word speech recognition apparatus of the present invention. In the figure, reference numeral 111 denotes a microphone which performs voice input and converts voice information into an electric signal;
Reference numeral 112 denotes an amplifier for amplifying an audio signal converted to an electric signal; 113, an A / D converter; 114, an operating system (OS) 1141;
142, representative distribution 1143, semi-continuous HMM 1144
A memory for storing the word dictionary 1145, the work area 1146, etc., 115 is an arithmetic processor (CPU), 116
Are other peripheral devices such as a printer and a display device. The microphone 111 of FIG. 11 is provided to the voice input unit 1 of FIG. 3 and the voice analysis unit 2, the voice detection unit 3, and the sound verification unit 4 (probability calculation unit 41, probability mixing unit 42, probability accumulation unit 4) of FIG.
3), standard pattern storage means 5 (representative distribution storage means 51,
The functions of the semi-continuous HMM storage means 52), the word evaluation means 6, the word dictionary 7, and the determination means 8 are realized by programs and various data stored in the arithmetic processor 115 and the memory 114 in FIG.

Also, the voice analysis means 2, the voice detection means 3, the sound collation means 4 (probability calculation means 41, the probability mixing means 42, the probability accumulation means 43), the standard pattern storage means 5 (the representative distribution storage means 51) in FIG. , The semi-continuous HMM storage means 52), the word evaluation means 6, the word dictionary 7, and the determination means 8, ie, the arithmetic processor 115 in FIG.
And a speech recognition program / representative distribution of the memory 114, a semi-continuous HMM, a word dictionary, and the like can be incorporated on a semiconductor chip to realize a microcomputer device for word speech recognition.
(Parsonal Digital Assistant) can be incorporated into various information devices that require voice recognition, and the application range is wide.

As described above, in the above embodiment,
For simplicity, we explained the case of word speech recognition,
In addition to similar syllable-based standard patterns and word dictionaries, the grammar that describes the sentence to be recognized as a sequence of words is stored. Matching is performed for each standard pattern to obtain a matching result in time series, and in the evaluation means, each sentence is evaluated based on the matching result for each standard pattern obtained as a time series and the word dictionary and the grammar information, By obtaining a recognition result according to the result, a sentence speech recognition apparatus, a microcomputer device for sentence speech recognition, and a sentence speech recognition method for recognizing sentence speech such as continuously uttered sentences and conversation sentences are realized. It is also possible.

Further, if the steps constituting the above-mentioned word speech recognition method and sentence speech recognition method are converted into program codes and recorded on a recording medium such as a CD-ROM or FD (flexible disk), they can be easily distributed to the market. The speech recognition method of the present invention can be widely spread.

According to the above embodiment, the intended object of the present invention is as follows:
In other words, the final evaluation result is obtained for all the hypotheses to be recognized, and the processing amount is small and the recognition amount is small so that the amount of sound matching processing is not proportional to the number of words to be recognized and can be suppressed by a certain amount. It is possible to obtain a word-speech recognition device and a sentence-speech recognition device with less deterioration in accuracy, a microcomputer device therefor, a speech recognition method for words and sentences, and a computer-readable recording medium recording the recognition method.

[0035]

As described above, according to the present invention, the number of probability calculations required for speech recognition can be greatly reduced, and large vocabulary speech recognition with a small processing amount can be performed while maintaining recognition accuracy.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of a speech recognition device of the present invention.

FIG. 2 is a diagram illustrating a hidden Markov model (HMM) of a basic recognition unit used in the speech recognition apparatus of the present invention.

FIG. 3 is a block diagram showing a detailed configuration of one embodiment of the speech recognition device of the present invention.

FIG. 4 is a diagram illustrating a representative distribution storage unit according to the present invention.

FIG. 5 is a diagram illustrating a semi-continuous Hidden Markov Model (HMM) used in the speech recognition apparatus of the present invention.

FIG. 6 is a diagram illustrating a representative distribution probability holding unit according to the present invention.

FIG. 7 is a diagram illustrating a semi-continuous Hidden Markov Model (HMM) used in the speech recognition device of the present invention.

FIG. 8 is a diagram illustrating a time series of a sound matching result.

FIG. 9 is a flowchart illustrating a word evaluation calculation process in a word evaluation unit.

FIG. 10 is a diagram illustrating an image of a word evaluation calculation process in a word evaluation unit.

FIG. 11 is a block diagram illustrating a specific hardware configuration of the speech recognition device in FIG. 3;

[Explanation of symbols]

1: voice input means, 2: voice analysis means, 3: voice detection means, 4: sound collation means, 5: standard pattern storage means, 6:
Word evaluation means, 7: word dictionary, 8: judgment means, 41: probability calculation means, 42: probability mixing means, 51: representative distribution storage means, 52: semi-continuous HMM storage means, 111: microphone, 112: amplifier, 113: A / D converter, 114:
Memory, 1141: operating system (O
S), 1142: speech recognition program, 1143: representative distribution, 1144: semi-continuous HMM, 1145: word dictionary, 1146: work area, 115: arithmetic processor (CPU), 116: other peripheral devices.

Claims (26)

[Claims] 1. A voice input unit for inputting a voice, a voice analysis unit for analyzing an input voice (input voice) and outputting a time series of feature vectors, and a standard pattern for a basic voice unit are stored. A standard pattern storage unit, a word dictionary that describes words of a recognition target word as a sequence of basic speech units, a matching unit that matches a time series of a feature vector of the input speech with the standard pattern, based on the matching result An evaluation means for evaluating the recognition target by using the comparison means, wherein the collation means collates each of the standard patterns over the entire section of the feature vector time series of the input speech and a collation result for each standard pattern. As a time series, wherein the evaluation means evaluates each word based on the matching result for each standard pattern obtained as the time series and the information of the word dictionary. And a means for obtaining a recognition result according to the evaluation result. 2. A speech analysis means for analyzing input speech (input speech) and outputting a time series of feature vectors, a standard pattern storage means for storing a standard pattern for a basic speech unit, and a recognition target word. A word dictionary that describes the words as a sequence of speech basic units, a matching unit that matches the time series of the feature vector of the input speech with the standard pattern, and an evaluation unit that evaluates a recognition target based on the matching result. Wherein the matching means matches each of the standard patterns over the entire section of the feature vector time series of the input speech, and obtains a matching result as a time series for each standard pattern. The evaluation unit evaluates each word based on the result of the comparison for each standard pattern obtained as the time series and the information of the word dictionary, and Accordingly, a microcomputer device for word speech recognition, which is a means for obtaining a recognition result. 3. A voice input unit for inputting voice, a voice analysis unit for analyzing the input voice and outputting a time series of feature vectors, and a standard pattern storage unit for storing a standard pattern for a basic voice unit. And a word dictionary that describes the words of the recognition target word as a sequence of basic speech units, a grammar that describes the recognition target sentence as a sequence of words, and matches the time series of the feature vector of the input voice with the standard pattern. And a evaluating unit that evaluates a recognition target based on the matching result, wherein the matching unit determines each of the standard patterns in the entire interval of the feature vector time series of the input speech. Means for obtaining a matching result as a time series for each of the standard patterns, wherein the evaluation means includes a matching result for each of the standard patterns obtained as the time series and the word dictionary. A sentence speech recognition apparatus characterized in that it is means for evaluating each sentence on the basis of information on a book and the grammar, and obtaining a recognition result according to the evaluation result. 4. Speech analysis means for analyzing input speech and outputting a time series of feature vectors, standard pattern storage means for storing a standard pattern for a speech basic unit,
A word dictionary that describes the words of the recognition target word as a sequence of basic speech units, a grammar that describes the recognition target sentence as a sequence of words, and a collation that matches the time series of the feature vector of the input speech with the standard pattern Means, and a sentence speech recognition microcomputer device having evaluation means for evaluating a recognition target based on the matching result, wherein the matching means converts each of the standard patterns into an entire section of the feature vector time series of the input speech. Means for obtaining a matching result as a time series for each standard pattern, wherein the evaluation means is based on the matching result for each standard pattern obtained as the time series and the information of the word dictionary and the grammar. A means for evaluating each sentence and obtaining a recognition result according to the evaluation result. 5. A voice input step of inputting voice, a voice analysis step of analyzing input voice (input voice) and outputting a time series of feature vectors, a time series of feature vectors of the input voice, A word-speech recognition method comprising: a matching step of matching a standard pattern for a basic voice unit; and an evaluating step of evaluating a recognition target based on the matching result. A step of obtaining a matching result for each standard pattern as a time series by performing matching over the entire section of the speech feature vector time series, and the evaluation step includes the step of obtaining a matching result for each standard pattern obtained as the time series and recognizing the same. Each word is evaluated based on information of a word dictionary that describes the words of the target word as a sequence of basic speech units, and a recognition result is obtained according to the evaluation result. Word speech recognition method which is a Mel step. 6. The standard pattern is constituted by an appearance probability distribution of a feature vector of the speech, and the matching step includes calculating a probability of each standard pattern from the feature vector of the input speech and the appearance probability distribution. 6. The word speech recognition method according to claim 5, wherein the matching is performed based on the calculated probability value. 7. The word speech recognition method according to claim 5, wherein the matching step performs a matching calculation based on a dynamic programming. 8. The word speech recognition apparatus and word speech recognition method according to claim 7, wherein said dynamic programming method uses a Viterbi algorithm. 9. The word speech recognition method according to claim 5, wherein the basic speech unit is a syllable. 10. The word speech recognition method according to claim 5, wherein the basic speech unit is a triphone chain of vowels, consonants, and vowels. 11. The matching step includes matching each of the standard patterns over the entire section of the feature vector time series of the input voice and obtaining a matching result as a time series for each of the standard patterns. The word speech recognition method according to any one of claims 5 to 10, wherein only a portion where the middle evaluation result is higher than a predetermined reference value is obtained as a time series of the matching result. 12. The collating step comprises: collating each of the standard patterns over the entire section of the feature vector time series of the input voice to obtain a collation result as a time series for each standard pattern; 11. The method according to claim 5, wherein only the portion where the middle evaluation result is maximum in the time direction is obtained as a time series of the comparison result.
The word speech recognition method according to any one of the above. 13. The collation result obtained as a time series for each standard pattern includes an evaluation value of each standard pattern ending at that time for each time and corresponding start point information. 13. The word speech recognition method according to claim 11, wherein 14. The evaluation method according to claim 5, wherein the evaluation step terminates the evaluation of a recognition target less than a predetermined reference value during the evaluation. The described word speech recognition method. 15. A computer-readable recording medium recording each step of the word speech recognition method according to claim 5. Description: 16. A voice input step of inputting voice,
A voice analysis step of analyzing an input voice (input voice) and outputting a time series of a feature vector; a matching step of matching the time series of the feature vector of the input voice with a standard pattern for a basic voice unit; An evaluation step of evaluating a recognition target based on the collation result, wherein the collation step collates each of the standard patterns over the entire section of the feature vector time series of the input speech, and executes each standard pattern. A step of obtaining a collation result as a time series for each of the words, wherein the evaluation step comprises: a word dictionary that describes the collation result and the word of the recognition target word for each standard pattern obtained as the time series as a sequence of speech basic units; Each sentence is evaluated based on grammatical information describing the recognition target sentence as a sequence of words, and a recognition result is obtained according to the evaluation result. A sentence speech recognition method characterized by the following steps: 17. The standard pattern is constituted by an appearance probability distribution of a feature vector of the speech, and the matching step calculates a probability of each standard pattern from the feature vector of the input speech and the appearance probability distribution, 17. The sentence speech recognition method according to claim 16, wherein matching is performed based on the calculated probability value. 18. The sentence speech recognition method according to claim 17, wherein the matching step performs a matching calculation based on a dynamic programming. 19. The sentence speech recognition method according to claim 18, wherein said dynamic programming uses a Viterbi algorithm. 20. The sentence speech recognition method according to claim 16, wherein the basic speech unit is a syllable. 21. The speech basic unit according to claim 16, wherein the vowel, consonant, and vowel are a three-phoneme chain.
10. The sentence speech recognition method according to any one of items 9 to 9. 22. The collating step comprises: collating the standard patterns over the entire section of the feature vector time series of the input voice to obtain a collation result as a time series for each standard pattern; The sentence speech recognition method according to any one of claims 16 to 21, wherein only a portion where the middle evaluation result is higher than a predetermined reference value is obtained as a time series of the matching result. 23. The collating step comprises: collating each of the standard patterns over the entire section of the feature vector time series of the input speech to obtain a collation result as a time series for each standard pattern; The method according to claim 16, wherein only a portion where the middle evaluation result is maximum in the time direction is obtained as a time series of the matching result.
Item 1. The sentence speech recognition method according to any one of 1. 24. The collation result obtained as a time series for each standard pattern includes an evaluation value of each standard pattern ending at that time for each time and corresponding start point information. The sentence speech recognition method according to any one of claims 16 to 23. 25. The method according to claim 16, wherein in the evaluation step, the evaluation is terminated in the middle of the evaluation for a recognition target less than a predetermined reference value in the middle of the evaluation. The sentence speech recognition method described. 26. A computer-readable recording medium recording each step of the sentence speech recognition method according to claim 16. Description: