JP3575904B2 - Continuous speech recognition method and standard pattern training method - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a continuous speech recognition method and a standard pattern training method, more specifically, a pattern representing a part of a class is connected in the time direction to form a state transition model, and a matching duration time of each state in the state transition model. A continuous speech recognition method that obtains recognition results by comparing input speech patterns while comparing the state transition model with feature patterns of the input speech while controlling The present invention relates to a standard pattern training method suitable for training a standard pattern necessary for the training.
[0002]
[Prior art]
First, the symbols used in the present specification are defined as follows.
[0003]
[Outside 1]
[0004]
First, a conventional continuous speech recognition method will be described. Now, it is assumed that there are S standard pattern sequences for the input voice pattern, and the s-th sequence is (s) W. (S) W is composed of L standard patterns connected as shown in the following equation. This standard pattern is a pattern that characterizes a type of speech (for example, a phoneme or a word).
[0005]
(Equation 1)
[0006]
However, L is variable. Here, q (l) is the index of the l (1 ≦ l ≦ L) standard pattern in the sequence, and has V vocabulary numbers.
Similarly, the column X of the input speech feature is represented as follows.
X = ｛x₁, ..., x_m, ..., x_M…… (2)
Here, the problem of continuous speech recognition corresponds to finding the reference sequence * W that minimizes the distance D (X, (s) W) between the uttered speech X and the reference sequence.
[0007]
(Equation 2)
[0008]
The minimization on the right side of Equation (4) is performed on the number of connected standard patterns, the arrangement of models, and the matching function, respectively. Equation (4) can be obtained by dynamic programming. Here, θ is a function representing the collation path. To create the standard pattern sequence (s) w, Nakagawa, Hidden Markov Model (HMM), which is described in detail in the Institute of Electronics, Information and Communication Engineers (1988), a neural network, It is modeled by arithmetic averaging of voice patterns.
[0009]
A reference sequence is created based on the combination of the standard patterns Wi. If there is no restriction on the combination, the search space at the time of collation increases and the recognition performance decreases. Therefore, a language model is introduced to give various language constraints. For example, a language model based on syntax control is given in terms of 0, 1 and is described in a context-free grammar, etc., and is described in detail in LR (Left-to-right) described in ATR, "Automatic Translation Telephone" Ohmsha (1994). (Rightmost derivation) using a parser or the like. In the recognition method according to the above-mentioned document, whether to reject or continue the hypothesis obtained from the parser is determined based on the likelihood obtained from the phoneme HMM at the same time as the analysis. Finally, the hypothesis having the highest likelihood is set as the recognition result. In this case, Wi in equation (1) must be a sequence received by LR parsing and corresponding to the terminal symbol.
[0010]
Next, a conventional standard pattern training method will be described. For example, suppose that one wants to extract the word date / one day / one / from the utterance. There are various ways in which the speaker speaks, such as (1) / one day / month / continuous utterance, (2) / one month / one day / (_: slight pause section), (3) A word other than the recognition target may be inserted between words such as / one day of the month /. For such an utterance, creating all of the above three patterns in the standard pattern used for collation causes an increase in the pattern storage capacity. Therefore, short patterns such as / month /, / day / It is common to create standard patterns that use words as units. Such a standard pattern and input speech are collated by using a spotting method described in Nakagawa, "Speech Recognition by Probabilistic Model", IEICE (1988), and keywords are extracted. I do.
[0011]
Training of the standard pattern can be realized by uttering isolated words such as normal / month /, / day / several times and calculating the arithmetic average of the characteristic pattern. However, the standard pattern using such discretely uttered voices is different in form from the continuous uttered voices as described in (1) to (3) above. Therefore, the pause part (2) and the /// part of (3) that are not the recognition target are extracted as one of the target words to generate a spurt, or a pattern or utterance speed representing a word in continuous speech. Is different from that of an isolated word, and may be dropped even though it is the target word.
[0012]
The above phenomenon occurs because the standard pattern for the speech style is not precisely designed. To cope with this problem, there is a method using a garbage model as disclosed in JP-A-7-36479. In this method, a model corresponding to a word other than a registered word is created, and a standard pattern is trained so that utterance parts other than a keyword are absorbed by the model. Also, as described in the International Telecommunications Research Institute, edited by "Automated Translation Telephone" Ohmsha (1994), all phenomena that can occur as utterances are described in context-free grammar, etc. There is a method of matching a hidden Markov model (HMM: Hidden Markov Model) in units of phonemes with an input voice using a generalized LR (Left-to-right Rightmost derivation) analysis algorithm.
[0013]
[Problems to be solved by the invention]
Recently, in the creation of a standard pattern in the conventional continuous speech recognition method described above, Takami et al., “Automatic Generation of Hidden Markov Network by Sequential State Division Method”, IEICE Transactions, Vol. J76-D-II, No. 10, pp. As reported in 2155-2164 (1993-3), a model has been proposed that considers not the phoneme unit but the environment of the phoneme. For example, when recognizing / k / of a voice uttered as / aka /, matching is performed using an HMM / a-ka-a / having information that / a / is present before and after / k /. Do. Similarly, when / ik / is uttered, / k / uses the HMM of / ik-i /. In both of the above utterances, the consonant at the center is / k /, and the same model / k / is used for collation in the case of the phoneme environment independent type, but in the case of the environment dependent type, A different HMM will be used. Therefore, at the stage of designing a phoneme model, not only the phoneme but also the movement path from a certain phoneme to a phoneme in the acoustic space can be modeled, and high-accuracy recognition performance can be expected.
[0014]
On the other hand, various LR parsers have been proposed to drive such a phoneme environment-dependent model as a verifier. Nagai et al., "Continuous Speech Recognition Integrating Hidden Markov Network and Generalized LR Parsing", IEICE Transactions, Vol. J77-D-II, No. 1, pp. 9-19 (1994-1) reports an example in which the analysis algorithm is changed to a phoneme environment dependent type using a phoneme environment independent type LR table. In this example, the phoneme environment becomes independentBut driveDedicatedPhoneme environment dependentAn analyzer must be developed.
[0015]
Nagai et al., "Conversion Algorithm from Context-Free Grammar to Phoneme Context-Dependent Grammar", Proc. 81-82 (1992-3), a method for converting a phoneme environment-independent LR table into a phoneme environment-dependent LR table, and a method for converting a phoneme environment-independent context-free grammar into a phoneme environment-dependent context-free grammar. It introduces how to convert to grammar. However, these methods attempt to convert a phoneme environment-independent LR table or a context-free grammar into a phoneme environment-dependent type, assuming a general-purpose task. It is expected that the number will explode.
[0016]
The present invention has been made in view of the actual situation of the conventional continuous speech recognition method as described above, generates a context-free grammar considering the environment of speech and the like according to a task, has a simple mechanism, and has a small storage amount. The purpose of the present invention is to provide a continuous speech recognition system that can perform high-speed and high-accuracy matching by narrowing the matching range of speech recognition using a syntax analysis unit and adaptively training standard patterns that take the environment into consideration. It was done as.
[0017]
In the conventional standard pattern training method described above, in the method using a garbage model, in order to design a model other than a registered word as a relatively coarse model, words to be extracted are also attracted to the garbage model and absorbed. May be lost. Therefore, the model parameters must be carefully controlled. It is also conceivable to increase the number of garbage models in order to avoid unnecessary absorption, but the storage capacity of the models increases.
[0018]
On the other hand, in the method using the predictive generalized LR algorithm, post-processing for checking whether or not a keyword is present in the recognition result is required in order to recognize every single phrase of the utterance content. Further, the number of grammatical rules for dealing with speech phenomena increases, and the description becomes complicated, so that management is not easy.
[0019]
Therefore, the present invention has been made in view of the actual situation of the conventional standard pattern training method as described above, and uses a standard pattern group having a small storage amount and an LR table having a simple mechanism and a small storage amount.StructureBy selecting a standard pattern directly from the syntax analysis unit by the sentence analysis unit, the standard pattern training method that improves the training efficiency of the standard pattern and the recognition accuracy for the utterance style and enables highly accurate keyword recognition in a short time It was made for the purpose of providing.
[0020]
[Means for Solving the Problems]
The invention according to claim 1 includes means for extracting a characteristic amount of an input voice, means for connecting a pattern representing a part of a class in a time direction to form a state transition model, and modeling means for a voice class; Parsing for parsing symbol strings by grammarDepartmentAnd means for matching an input voice pattern while controlling the matching duration time of each state in the state transition model, and comparing the state transition model with the feature pattern of the input voice to obtain a continuous speech. In recognition method, SaidUsing the phonetic symbol sequence received by the parsing unit, a terminal symbol sequence including the surrounding environment of the class is generated, and the grammar is created. Collate.
[0021]
The invention according to claim 2 is the invention according to claim 1,SaidA speech uttered based on the speech symbol string received by the syntax analysis unit is used as an input, and a state transition model including a surrounding environment of a kind corresponding to the input is connected and trained.
The invention according to claim 3 is the invention according to claim 1,SaidBy inputting the speech including the phonetic symbol string received by the syntax analysis unit, collating it with the state transition model including the environment before and after the class, displaying a predetermined number of the recognition results in plausible order, and selecting the correct candidate And train the correct state transition model.
The invention of claim 4 relates to the training of the state transition model including the kind of surrounding environment according to the invention of claim 2 or 3, wherein if the training has been performed on the model in the past, the past state transition is performed. Superimpose with modelTrainYou.
[0022]
The invention of claim 5 relates to the training of the state transition model including the kind of surrounding environment in the invention of claim 2 or 3, and if the training has been performed on the model in the past, the training is newly performed. modelKind ofToGenerate a state transition model including the environment before and afterTo train andResponseState transition modelDoes not train.
The invention of claim 6 relates to the training of a state transition model including the kind of surrounding environment according to the invention of claim 4, and, among the state transition models stored by claim 5, a model most similar to the input voice is determined. Update.
The invention of claim 7 relates to the training of a state transition model including the kind of before and after environment in any one of the inventions of claims 2 to 6, and connects an environment independent state transition model as an initial model.
[0025]
BEST MODE FOR CARRYING OUT THE INVENTION
First, the continuous speech recognition method will be described.
FIG. 1 is a schematic block diagram for explaining an embodiment of a continuous speech recognition system according to the present invention. In the figure, 1 is an LPC analysis unit, 2 is a collation unit, 3 is an environment-dependent grammar unit, and 4 is an environment-dependent grammar unit. Operation table part, 5syntaxAn analysis unit, 6 is an environment-dependent DST model, 7 is a pattern connection unit, 8 is a judgment unit, 9 is a switch, 10 is an environment-independent grammar unit, 11 is an environment-independent operation table unit, and 12 is a symbol processing unit. According to the embodiment shown insyntaxThe environment-dependent DST model 6 can be used without changing the analysis unit 5, and adaptive and more reliable recognition of the task can be performed. The environment-independent grammar unit 10 stores a grammar having a normal phoneme as a terminal symbol using a context-free grammar or the like. Table 1 shows examples of the grammar. In Table 1, lowercase letters on the right side represent terminal symbols. In the present embodiment, the grammar terminator and the standard pattern are used as phonemes, but a class such as a word or a syllable may be employed. The LR analysis table obtained from the grammar is stored in the environment independent operation table section 11. The contents of Table 1 are the same as the LR analysis table described in detail in AV Aho et al., "Compilers-Principles, Techniques, and Tools", Addison-Wesley (1986), etc., from the ACTION section and the GOTO section. It is made up.
[0026]
[Table 1]
[0027]
First, switch 9 is set to the A side to create a phoneme environment-dependent grammar., StructureThe sentence analysis unit 5 is driven to output an acceptable sentence using the terminal symbol string. This is described in detail in Kita et al., "Continuous Speech Recognition Using HMM Phoneme Recognition and Extended LR Parsing", Transactions of Information Processing Society of Japan, Vol. 31, 3, pp. 472-480 (1990). Thus, while predicting the terminal symbol to be analyzed next from the operation table,syntaxThis can be realized by driving the analysis unit 5.
[0028]
From the obtained sentences, the symbol processing unit 12 selects a sentence required as a recognition task. For the selection, a required sentence may be automatically selected by symbol string collation, or may be selected by a human editing an output result. Thereafter, the selected sentence is converted into an environment-dependent terminal symbol sequence according to the sequence of terminal symbols. For example, when the sentence "/ koreokure /" is obtained, the conversion is performed as "/ -ko kore reo eok oku ku ure re- /" in consideration of one character preceding and succeeding the target symbol. The symbol at the center is the target terminal symbol, and a symbol indicating the environment is added to the left and right. In the above example, / kor / indicates that a terminal symbol of / o / is preceded by a terminal symbol of / k /, and a symbol of / r / follows. // indicates the beginning or end of the symbol. In this embodiment, the number of preceding and succeeding symbols is one, but any number may be set. Next, an environment-dependent grammar is created using the converted terminal symbols and stored in the grammar section. Table 2 shows the grammar created. The LR analysis table obtained from the grammar is stored in the environment-dependent operation table section 4.
[0029]
[Table 2]
[0030]
Next, the switch 9 is turned to the B side to perform continuous voice recognition. The input voice is subjected to LPC analysis to extract 10-dimensional cepstrum parameters. However, the analysis conditions are a sampling frequency of 8 kHz, windowing with a Hamming window (window width 16 ms), and an LPC analysis order of 14. The shift width per frame is set to 5 msec. The analysis method is not limited to the above, and as described in detail in Niimi, "Speech Recognition", Kyoritsu Shuppan (1979), any acoustic analysis method such as frequency analysis can be used. Good.
[0031]
syntaxThe analysis unit 5 determines which phoneme to match from the LR analysis table. SolutionAnalysisEach time the state progresses, the continuation as detailed in Muroi et al., "Word Speech Recognition Using Duration Control State Transition Model", J72-D-II, 11, pp.1769-1777 (1989-11) A time-controlled state transition (DST) model is connected. In the present embodiment, a DST model that takes into account the phoneme environment is used, and the matching unit compares the DST model with the feature amount of the input speech. The parsed phrase structure issyntaxAnalysis section5Record on the chart. Finally, the candidate having the lowest score among the candidates for which all analysis has been completed is obtained according to the equation (5), and is output as a recognition result.
[0032]
(Equation 3)
[0033]
Here, r is an expansion / contraction function obtained by the dynamic programming. With this function, the input feature value of the m-th frame to be collated is associated with the state of the r (m) -th DST model. l_{r (m)}Indicates the frame length in the r (m) -th partial pattern when the input voice pattern is divided into N (s) partial patterns. The first item on the right side represents the distance related to the feature amount obtained by the acoustic analysis, and the second item represents the distance related to the duration of the partial pattern. a is a positive number and determines how much the distance related to the duration is reflected in the overall distance. In this embodiment, a is set to about 0.1. By using the above-described DST model, it is possible to perform matching in consideration of not only a feature amount in an acoustic space but also a special structure of a voice pattern (particularly, a time length of a partial pattern).
[0034]
FIG. 2 is a schematic block diagram for explaining another embodiment of the present invention, in which 13 is an utterance list, 14 is a DST model training unit, and other operations are the same as those of the embodiment shown in FIG. Are given the same reference numerals as in FIG. Thus, the embodiment shown in FIG. 2 enables training of a phoneme environment-dependent DST model using the environment-dependent grammar and the operation table obtained by the embodiment shown in FIG. First, the switch 9 is turned to the A side, and the training of the phoneme environment-dependent DST model 6 is performed. A voice corresponding to the utterance list 13 is input, and LPC cepstrum parameters are extracted. Next, in accordance with the utterance list 13, the sequence is collated with the environment-dependent DST model sequence using a dynamic programming method, and the expansion function θ is minimized according to the criterion of Expression (4). The obtained expansion function is defined as r.
The DST model training unit 14 updates the average value and the duration of the model according to the following equation. Where N_{r (m)}Is the last frame number of the input pattern associated with the r (m) -th state of the DST model.
[0035]
(Equation 4)
[0036]
Where N_{r (0)}= 0.
After performing the above training, the switch 9 is turned to the B side, and continuous voice recognition is performed. The configuration of the recognition process is the same as that of the embodiment of FIG.
[0037]
FIG. 3 is a schematic block diagram for explaining still another embodiment of the present invention, in which 15 is a result display section, 16 is a selection section, and other than the embodiment shown in FIG. 1 or FIG. Parts that perform a similar function are given the same reference numerals as in FIG. 1 or FIG. Thus, in the embodiment shown in FIG. 3, a phoneme environment-dependent DST model can be trained using input speech uttered for recognition. According to the embodiment shown in FIG. 3, recognition and DST model training can be performed simultaneously. First, an input voice is recognized in the same process as in the embodiment of FIG. When the correct answer is included in the display unit 15, the correct answer can be selected by the selecting unit 16 such as a keyboard. With this selection, it is possible to determine a DST model sequence to be trained for the input speech pattern. Equations (4), (8), and (9) are applied to these DST model strings, and the training unit 14 updates the average value and the duration of the DST model. The training process is the same as the embodiment of FIG.
[0038]
In this embodiment, since the Euclidean distance shown in Expression (7) is used as the distance scale at the time of collation on the display unit, candidates are displayed in the order of the lowest score. If scores of recognition candidates are given based on likelihood or the like, they are displayed in descending order of the scores. Of course, either criterion may be used in the present invention.
[0039]
In the DST model training unit 14 of FIG. 2 or FIG. 3, when a model trained in the past exists for the same class, the DST model is trained by the following two methods. One is that the previously trained model W_k1And the newly trained model W_k2And W_n3How to create
W_k3= BW_k1+ (1-b) W_k2          … (10)
Here, b is a positive number indicating the mixture ratio between the past model and the new model. As a special case, b = 0 indicates that the model is not trained, and b = 1 corresponds to replacing with a new model.
Another method is to store both the old model and the new model. That is, each time a training voice is input, a new DST model is created. At the time of recognition, a DST model sequence closest to the input speech pattern may be output as a recognition result.
[0040]
Further, a method combining the above two training methods is also possible. In the second method described above, the recognition accuracy can be improved by having a plurality of models for the same class, but the recognition time becomes longer because the number of combinations at the time of matching increases. Therefore, after a predetermined number of models have been created, a model to be superimposed is selected, and the selected model and the newly trained model are superimposed according to Equation 10. If the column s is a column including the DST model to be superimposed, the DST model sequence to be superimposed is:
[0041]
(Equation 5)
[0042]
Meet. With this method, the relationship between the recognition time and the recognition accuracy can be freely adjusted, and the performance desired by the user can be set.
In order to train the environment-dependent DST model described above, an environment-independent DST model can be used as an initial model. For example, consider training a DST model of / a-ka-a / where the leading and succeeding phonemes are / a /. As an initial model in this case, consider training a phoneme environment independent DST model of / k /. As an initial model in this case, training is started using a phoneme environment independent DST model of / k /. By giving a good initial value from the phoneme environment independent DST model, a highly accurate model can be designed.
[0043]
Next, the standard pattern training method will be described.
Figure 4, MarkFIG. 2 is a schematic block diagram for explaining an embodiment of a quasi-pattern training method, in which 21 is a segmentation unit, 22 is a feature pattern creation unit, 23 is a matching unit, 24 is a cumulative score storage unit, and 25 is a comparison unit. , 26 is an LR table, and 27 is a predictive chart parsing unit.₁Into the A side, and train the standard pattern. In FIG. 4, a predictive chart parsing unit 27 using an LR table unit 26 is driven in order to create a state transition model for an input voice. The LR table 26 stores an operation table obtained from a grammar as shown in Table 3. In the symbols in Table 3, terminal symbols start with '*', and other symbols are non-terminal symbols. This description has been simplified to illustrate the example, but is context free.SentenceA more complex description is possible if it is a notation by the method.
[0044]
[Table 3]
[0045]
The contents of the LR table are as follows: V. It is the same as the LR analysis table described in detail in Aho et al., "Compilers-Principles, Techniques, and Tools", Addison-Wesley (1986), and is composed of an ACTION section and a GOTO section. There are four types of operations in this table: state transition, application of grammar, acceptance, and error.
[0046]
Using the LR table in Table 3, the predictive chart parsing unit 27 extracts terminal symbols one by one from the beginning, applies the algorithms shown in Tables 4 to 6, and records the results as a chart shown in Table 7. I do. All phrase structures are recorded on the chart until the receiving operation is finally performed. However, '*' is set at the last position of the terminal symbol sequence predicted by the terminal symbol representing the end.
[0047]
[Table 4]
[0048]
[Table 5]
[0049]
[Table 6]
[0050]
Table 7 shows the analysis result of "1:00 on January 1" as an example. In addition, based on the grammar, "2:00 on January 1", "1:00 on January 2" etc. are sequentially displayed. Generated. Training of the standard pattern can be realized by creating a state transition model corresponding to these symbol sequences.
[0051]
[Table 7]
[0052]
By the operation of the predictive chart parsing unit 27 described above, the index numbers of the character sequences constituting the terminal symbols are sequentially sent to the standard pattern storage unit 28. Since the standard patterns are stored in units of characters, the connecting unit 29 refers to the index numbers and connects the standard patterns in units of terminal symbols, and the state transition model unit 30 creates a state transition model. For example, if the standard pattern is stored in phoneme units, the terminal pattern is composed of the standard pattern of / i, ch, i, g, a, t, u / with respect to the terminal symbol "January". The state transition model may be expressed by a probability model such as an HMM, or may be strictly expressed by a word graph or a finite state network.
[0053]
On the other hand, the input speech is input by the segmentation unit 21 for a predetermined time, and is converted into a feature pattern by an analysis method as described in detail in Niimi, "Speech Recognition", Kyoritsu Shuppan (1979) and the like. . Here, 10-dimensional cepstrum parameters are extracted and used as feature patterns. However, the analysis conditions are as follows: sampling frequency: 16 kHz, high-frequency emphasis: first-order difference, 256-point Hamming window, update cycle: 10 ms, and LPC analysis order: 20. The analysis method is not limited to the above, and any acoustic analysis method such as frequency analysis may be used. The input speech includes a keyword corresponding to the terminal symbol generated from the chart parsing unit.
[0054]
Next, the state transition model created as described above and the feature pattern of the input voice are collated by the collation unit 23. The state transition model corresponding to the s-th terminal symbol in the terminal symbol string generated from the syntax analysis unit 27 is represented by sW, (s = 1,..., S). sW is composed of L standard patterns.
[0055]
(Equation 6)
[0056]
Where p_{q (l)}Is an index of the standard feature pattern corresponding to the l (1 ≦ l ≦ L) -th in the sequence, and has V standard patterns in total. Taking Table 7 as an example, since the number of terminal symbols of the generated sentence is 3, S = 3. Further, in the case of the embodiment, each standard pattern corresponds to a phoneme, so that the number of standard patterns is equal to the total phoneme number.
Similarly, the input feature pattern X is represented as follows.
X = ｛x₁, ..., x_m, ..., x_M｝… (13)
In the embodiment, X is a feature pattern including S keywords in the input voice. The matching unit obtains a matching score D between the input speech feature pattern and the state transition model by the following equation.
[0057]
(Equation 7)
[0058]
Where m_s1, M_s2Are the end points of the extraction section of the audio feature pattern corresponding to the s-th keyword, and represent the start point and the end point, respectively. The matching function r is a function representing a matching path, and can be obtained by a well-known dynamic programming method or the like. The matching function associates the input feature value of the m-th frame with the r (m) -th standard pattern constituting the keyword. Score D (x) between the standard pattern and the voice feature pattern_m, P_{r (m)}) Is obtained by subtracting the well-known Euclidean distance from a threshold value having a positive value. Since a partial pattern of the voice feature pattern corresponding to the standard pattern is obtained from r obtained by the equation (15), the standard pattern is trained using this partial pattern. This training is performed by calculating an arithmetic mean of the feature amount of the standard pattern and the feature amount of the partial pattern, and newly registering it as a standard pattern.
[0059]
The training method is not limited to the above, and if the state transition model is represented by HMM, D (x_m, P_{r (m)}) As the likelihood. The training of the HMM at this time can be performed by the Baum-Welch estimation method described in detail in Nakagawa, "Speech Recognition by Stochastic Model", and the like. Equation (15) is based on maximization, but is not limited to this, and the essence of the present invention does not change even if training is performed based on minimization criteria based on a simple Euclidean distance.
[0060]
As explained above,-Since the partial patterns of the input voice are collated in units of words, even if pauses or unnecessary words are inserted between keywords, training of standard patterns is possible. The above processingPredictive typeChart parser27The training is completed by repeating until no terminal symbol sequence is generated. Next, switch W₁In B, keyword recognition can be performed. At recognition,Predictive typechartsyntaxIt works so as to predict a keyword from the analysis unit 27 and the LR table unit 26. SolutionAnalysisEach time the process proceeds, a standard pattern is linked to create a state transition model of the predicted keyword. The collation unit 23 collates the state transition model with the feature amount of the input speech. The predicted score of the keyword candidate is stored in the cumulative score storage unit 24, and finally the candidate having the highest score among the candidates for which all the analysis has been completed is obtained according to the equation (15), and is output as the recognition result. I do.
[0061]
In the embodiment shown in FIG. 4, the collation time can be shortened by introducing the following condition in Expression (16).
[0062]
(Equation 8)
[0063]
This equation (16) indicates that the s-th keyword is detected in the input feature pattern, and the matching is started with respect to the state transition model of the next keyword, that is, the s + 1-th keyword, from a frame that falls within the section. ing.
[0064]
When a plurality of keywords are input, high-speed collation can be performed by cutting out only a portion where a voice exists in the segmentation unit. FIG. 8 shows a speech waveform including two keywords: / month / and / day /. As can be seen from FIG. 8, there is a slight pause between / month // day /. In such a case, the segmentation unit 21 uses the speech extraction algorithm described in Niimi, “Speech Recognition”, Kyoritsu Shuppan (1979), and the like, and sets the section between A and B in FIG. Ask for. After that, only the section between A and B that has been cut out is to be compared with the state transition model, so that the matching section can be shortened.
[0065]
Figure 5,other4 is a schematic block diagram showing an embodiment of the present invention, in which parts having the same functions as those of the embodiment shown in FIG. 4 are denoted by the same reference numerals as in FIG. In the embodiment shown in FIG. 5, a plurality of pairs of the LR table section (26A, 26B, 26C) and the predictive chart parsing section (27A, 27B, 27C) are prepared. If the standard pattern is trained from the initial stage using a voice including a plurality of keywords during the training of the standard pattern, the pattern may become unstable. In order to avoid such a phenomenon, at the initial stage, only a single keyword is trained from the input voice, and by gradually increasing the number of keywords included in the voice, not only the standard pattern is stabilized, but also the input voice A variety of speaking styles can be trained together. In the embodiment, the LR table 26A and the prediction typechartA sentence including one keyword is generated using the syntax analysis unit 27A. Similarly, the remaining two sets generate a sentence containing two keywords and a sentence containing three keywords, respectively. During training, ie, switch W₁When the switch is put into A, first, the switch W_TwoIs input to C, and a standard pattern is trained from the input voice including one keyword according to the same procedure as in the above embodiment. Next, switch W_TwoIs sequentially switched to D and E, the number of keywords included in the voice can be increased, and the standard pattern can be trained. At the time of keyword recognition, switch W₁In B, and all the switches W2 in C, D, and E. Since all predictable keyword candidates can be generated, the candidate having the highest score among them can be output as a recognition result.
[0066]
FIG. 6 is a schematic block diagram for explaining another embodiment. The embodiment shown in FIG. 6 is obtained by adding the display device 32 to the embodiment shown in FIG. Switch W during training₁Into the A side, and switch W₃Into C. On the display device 32, a sentence including the keyword generated by the predictive chart syntax analysis unit 27 is generated, and displayed on the display device 32 as "January 1". The speaker inputs a voice while watching this display. By performing the subsequent processing in the same manner as in the method described in the embodiment of FIG. 4, the training of the standard pattern is completed. Recognition, switch W₁Into the B side and switch W₃To C.
[0067]
FIG. 7 is a schematic block diagram for explaining still another embodiment. The embodiment shown in FIG. 7 is obtained by adding the conversion unit 33 according to the embodiment of FIG. In order to perform the reading conversion, the grammar for creating the LR table is changed as shown in Table 8. In Table 8, the reading corresponding to the date corresponding to the keyword is added as a rewriting rule. On the display device at the time of training, the read conversion unit 33 also displays the right side of the rewrite rule including the terminal symbol. As a result, it is possible to display “January (one day), one day (one day)”, and it is not necessary to read one day as “one day”. Can be encouraged.
[0068]
[Table 8]
[0069]
【The invention's effect】
As is apparent from the above description, according to the present invention, it is possible to adaptively generate a context-free grammar considering a kind of environment according to a task. Further, it is possible to combine environment-dependent phoneme models without changing the conventional LR-Chart parsing unit having a simple mechanism and a small storage amount. In addition, it is possible to adaptively train a duration-controlled state model in consideration of a kind of environment. As a result, continuous speech recognition that performs high-precision and high-speed matching can be realized.
The invention according to claim 1 includes means for extracting a feature amount of an input voice, means for connecting a pattern representing a part of a class in a time direction to form a state transition model, and modeling a class of a voice. Parsing that analyzes phonetic symbol strings by grammarDepartmentAnd means for matching an input voice pattern while controlling the matching duration time of each state in the state transition model, and comparing the state transition model with the feature pattern of the input voice to obtain a continuous speech. In the recognition method,SaidBy using the phonetic symbol sequence produced by the parser, a terminal symbol sequence including the surrounding environment of the class is generated, and the grammar is created. Can be matched with
The invention according to claim 2 is the invention according to claim 1,SaidA speech uttered based on a speech symbol string received by the syntax analysis unit is used as an input, and a state transition model including a kind of surrounding environment of a kind corresponding to the input can be connected for training.
The invention according to claim 3 is the invention according to claim 1,SaidSpeech including a phonetic symbol string received by the parser is input, collated with a state transition model that includes the surrounding environment, and the recognition result is obtained.WhenBy displaying a predetermined number in a likely order and selecting a correct candidate, it is possible to connect and train a correct state transition model.
The invention according to claim 4 relates to the training of the state transition model including the kind of environment before and after according to claim 2 or 3, wherein if the model has been trained in the past, the past state transition model Can be superimposed.
The invention according to claim 5 relates to the training of the state transition model including the front and rear environment of the kind according to claim 2 or 3,Kind ofIf training has been conducted forGenerate a state transition model including the environment before and afterTo train andResponseState transition modelDoes not train, adjusts the relationship between recognition time and recognition accuracy freely, and sets the performance desired by the usercan do.
The invention according to claim 6 relates to training of the state transition model including the kind of surrounding environment in claim 4, and updating the model most similar to the input voice from the state transition models stored by claim 5. can do.
The invention according to claim 7 relates to the training of the state transition model including the kind of front and rear environment according to any one of claims 2 to 6, wherein an environment independent state transition model can be connected as an initial model..
[Brief description of the drawings]
FIG. 1 is a block diagram showing an embodiment of continuous speech recognition according to one embodiment of the present invention.
FIG. 2 is a schematic block diagram for explaining another embodiment of the present invention.
FIG. 3 is a schematic block diagram for explaining another embodiment of the present invention.
FIG. 4MarkIt is a block diagram showing one example of quasi-pattern training.
FIG. 5otherFIG. 3 is a schematic block diagram illustrating an example of FIG.
FIG. 6otherFIG. 3 is a schematic block diagram illustrating an example of FIG.
FIG. 7SaFIG. 11 is a schematic block diagram showing still another embodiment.
FIG. 8Shows a speech waveform containing two keywords / Jan /// Day /FIG.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... LPC analysis part, 2 ... collation part, 3 ... Environment-dependent grammar part, 4 ... Environment-dependent operation table part, 5 ...ParsingUnit, 6 environment dependent DST model, 7 pattern connection unit, 8 decision unit, 9 switch, 10 environment independent grammar unit, 11 environment independent operation table unit, 12 symbol processing unit, 13 generation list , 14: DST model training unit, 15: result display unit, 16: selection unit, 21: segmentation unit, 22: feature pattern creation unit, 23: collation unit, 24: cumulative score storage unit, 25: comparison unit, 26 .. LR table, 27 predictive chart parsing unit, 28 standard pattern storage unit, 29 connecting unit, 30 state transition model unit, 31 training unit, 32 display unit, 33 reading conversion unit.

Claims (7)

A means for extracting the feature amount of the input speech, a pattern representing a part of the class is connected in the time direction to form a state transition model, a means for modeling the class of the speech, and a speech symbol string is analyzed by grammar. and parsing unit, by a means for collating the input voice pattern while controlling the collation duration of each state in the state transition model is compared with the characteristic pattern of the input sound voices with the state transition model, the recognition result In the continuous speech recognition method to obtain the above, by using the phonetic symbol string received by the syntax analysis unit, to generate a terminal symbol sequence including the environment before and after the class, by creating a grammar, including the environment before and after the class A continuous speech recognition method characterized by matching a state transition model with an unknown input speech. Claims before a speech uttered as an input based on the phonetic symbol string accepted by Ki構 sentence analyzing unit, characterized in that train coupled state transition model including a longitudinal environment of the kind corresponding to the input Item 2. The continuous speech recognition method according to Item 1. Before an input speech containing the phonetic symbol string is accepted by Ki構 sentence analyzing unit collates a state transition model including a longitudinal environment classes, and displays a predetermined number of the recognition result plausible sequentially select the correct candidate The continuous speech recognition method according to claim 1, wherein the training is performed by connecting the correct state transition models. Relates training state transition model including a longitudinal environment of the compound, the claims if trained with respect to the model was done in the past, which is characterized that you trained superimposed with a past state transition model 4. The continuous speech recognition method according to 2 or 3. It relates training state transition model including a longitudinal environment of the class, if the training with respect to the model was done in the past, produce a state transition model including a longitudinal environment against the class of new the model to perform training, continuous speech recognition system according to claim 2 or 3 past pairs 応状 state transition model is characterized in that no training. The method according to claim 4, wherein a model most similar to the input voice is updated from among the state transition models stored according to claim 5, with respect to training of the state transition model including the environment before and after the class. Continuous speech recognition method. 7. The continuous speech recognition method according to claim 2, wherein an environment-independent state transition model is connected as an initial model for training of the state transition model including the kind of surrounding environment.