JPH08248980A - Voice recognition device - Google Patents

Abstract

PURPOSE: To easily modify the contents of a dictionary without recompiling the syntax rules. CONSTITUTION: The device is provided with a syntax rules storage means 7 which stores the syntax rules, that sets the grammar for acceptable voices, in a compiled form, a dictionary storage means 5 which is independently constituted from the means 7 and at least records the acceptable words in terms of phoneme string inscription, output inscription and word classifications, a voice piece recognition means 2 which recognizes the inputted voices in terms of phoneme piece units, a grammar collating means 3 which refers to the recorded contents of the means 5 and 7 and collates whether the rows of the phoneme pieces are grammatically acceptable or not, an inputting means 12 which inputs word editing instructions for the means 5 from an external means and a dictionary control means 11 which modifies the recorded contents of the means 5 based on the instructions from the means 12. Thus, the contents of the dictionary are edited by the means 11 without attaching the syntax rules.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for recognizing uttered speech by referring to linguistic restrictions such as a dictionary and syntax rules, and in particular, from the outside without rewriting the syntax rules. The present invention relates to a voice recognition device capable of easily changing the contents of a dictionary according to the instruction.

[0002]

2. Description of the Related Art A voice recognition device for mechanically recognizing a voice uttered by a user with reference to linguistic restrictions such as grammars, etc. Has been applied to various fields such as inputting a document or inputting a search expression to a database system with a voice uttered in a unit of a command string or the like. Speech recognition is realized by roughly combining two processes, a phoneme unit recognition process and a language recognition process. The phoneme segment recognition process is a process of generating a symbol string corresponding to a sequence of phonemes, such as phonemes, syllables, and semisyllabic segments, from a waveform signal of speech,
The language recognition processing is a syntax analysis that analyzes how the symbol string output by the speech recognition processing is configured as a language, or a process that determines the statistical ease of occurrence of the symbol string,
This is a process of performing a semantic analysis or the like to analyze what meaning a symbol string has.

Here, in order to perform highly accurate continuous speech recognition, syntactic analysis for analyzing the structure of a word as a sentence such as the dependency relation between words is essential in language recognition processing. In such parsing, especially when recognizing a speech with a certain number or more of vocabulary such as hundreds of words, the number of symbol strings generated by the phoneme segment recognition process becomes large. Since it is necessary to handle a plurality of candidates and the number of rules such as dictionaries to be referred to increases, a huge amount of calculation processing must be performed. Therefore, in order to perform speech recognition in real time, it is necessary to perform a parsing method at high speed.

In order to perform high-speed parsing, it is effective to compile a grammar, which is a rule for parsing, in advance, so that the rule is a parser (syntax analyzer).
It becomes a format that is easy to refer to, and moreover, it becomes possible to efficiently refer to only the rules required at that time in the parsing process, and the parsing process can be performed at high speed. Japanese Unexamined Patent Publication No. 2-113297 discloses a method of recognizing speech by a combination of phoneme segment recognition processing and language recognition processing, in which parsing is performed by an LR parser using an LR table in which syntax rules are compiled. Is proposed. The LR table is composed of an action table in which syntax rules are compiled in advance and two action tables called a Go-to table. The LR parser sequentially reads the sentence to be analyzed from the beginning of the sentence and stores the analysis state according to the instructions of the LR table. Advance the analysis while changing the stack. (See Hirogo Nomura, "Basic Technology of Natural Language Processing," published by The Institute of Electronics, Information and Communication Engineers, Japan)

[0005]

In a voice recognition device in which a grammar such as a syntax rule is previously compiled as represented by an LR parser, even if a new word needs to be registered, this word registration is required. Is not easy and requires a long time for the registration process. That is, it is necessary to register a new word in the grammar before compiling and to compile the grammar again after that because the new word cannot be registered as it is in the grammar of the compiled form. The more complicated the grammar is, the longer it takes to compile this grammar, so it is impossible to restart speech recognition immediately after registering a new word.
There is a problem that the efficiency of the voice recognition processing is significantly reduced.

On the other hand, if the syntax analysis is performed without compiling the grammar, a new word can be registered quickly, but on the other hand, it takes a long time to process the syntax analysis. . In the first place, the purpose of using parsing is
This is to quickly realize speech recognition processing according to a complicated grammar, so the original purpose of realizing real-time processing of speech recognition if the parsing processing takes a long time is destroyed. I will end up. Therefore, in the past, while quickly realizing speech recognition according to a complicated grammar that requires parsing,
A voice recognition device has not been realized that can quickly perform a word editing process such as registration of a new word or deletion of a word that has already been registered.

The present invention has been made in view of the above conventional circumstances, and it is an object of the present invention to provide a speech recognition apparatus capable of quickly performing a syntactic analysis and an editing process to a dictionary for recording words. To aim. In particular, a voice recognition device according to a second aspect has an object to realize a quick dictionary edit process in a voice recognition device that performs an LR parser.
Further, in particular, the speech recognition apparatus according to claim 3 records a phoneme piece candidate at the beginning for each word classification such as a part-of-speech or a type of the word in a dictionary in advance, and based on the phoneme piece candidate, the phoneme piece candidate is stored. It is an object of the present invention to realize a quick dictionary editing process in a voice recognition device that performs one-sided recognition processing.

[0008]

In order to achieve the above object, the speech recognition apparatus according to claim 1 stores a syntax rule storing means for storing a syntax rule for defining an acceptable speech grammar in a compiled form. And a dictionary storage means that is configured independently of the syntax rule storage means, records acceptable words at least in the phoneme string notation, output notation, and word classification, and a phoneme that recognizes input speech in units of phonemes. A piece recognition means, a grammar collating means for collating whether the arrangement of phonemes is grammatically acceptable or not by referring to the recorded contents of the dictionary storing means and the syntax rule storing means, and an external word editing instruction to the dictionary storing means. It is characterized in that it comprises an input means for inputting from the outside and a dictionary management means for changing the recorded contents of the dictionary storage means based on an instruction from the input means.

According to a second aspect of the present invention, in the speech recognition device according to the first aspect, the grammar collating means corresponds to the recognized phoneme string sequence by referring to the recorded contents of the dictionary storing means. The word to be determined is determined, the LR parser performs grammatical matching with reference to the recorded content of the syntax rule storage means for the word, and the classification of the subsequent word is predicted based on the classification of the word, and the prediction is performed. It is characterized in that a process for determining a succeeding word is performed by using the classified word.

A speech recognition apparatus according to a third aspect of the present invention is the speech recognition apparatus according to the first aspect, in which the dictionary storage means has a phoneme segment candidate at the beginning recorded in advance for each word classification. The grammar matching means refers to the recorded content of the dictionary storage means for the recognized phoneme string sequence to determine the corresponding word, and refers to the recorded content of the syntax rule storage means for the word, and uses the LR parser. Performs grammatical matching and predicts the classification of the following word based on the classification of the word, predicts the succeeding phoneme piece from the recorded contents of the dictionary storage means based on the classification of the word, and the phoneme piece recognition means The input voice is recognized by finding the phoneme piece that becomes the predicted head phoneme.

Here, a phoneme piece refers to an acoustic unit such as a phoneme, a phoneme, a syllable, and a half syllable, and a word classification refers to a word part of speech, a word type, and the like. For example, when a search formula is input by voice in a database system, word types such as command names, extensions, item names, titles, and the like that make up the search formula correspond to word classification. Further, the dictionary storage means and the syntax rule storage means are not limited to being configured by separate storage devices in terms of hardware, but may be configured by logically dividing a storage device that is integrated in terms of hardware. It suffices that the dictionary management means can be accessed as mutually independent objects.

As a method of inputting voice from the outside to the voice recognition device, it is general to capture the voice of the user with a microphone. However, the voice from a remote place can be received via a communication path such as a telephone or a radio. It is also possible to enter. Also,
For phoneme piece recognition processing, collation processing using HMM (Hidden Markov Model) as a phoneme model is generally used because it has relatively high accuracy and can be processed at high speed, but DP matching, neural network, etc. The matching process of can also be used. Further, the dictionary storage means can be easily realized by expressing and recording the words in a tabular format, but if the dictionary storage means is configured as a dictionary having a trie dictionary part in which a phoneme sequence is arranged in a tree structure, the dictionary can be referred to. Can be performed efficiently and at high speed, and words can be registered and deleted by the dictionary management means more quickly.

The syntax rule can express a template, a finite state automaton, or the like, but the expression using a context-free grammar is preferable because the description is easy and the expressiveness is rich. Further, as an expression format of the compiled syntax rule stored in the syntax rule storage means, a method of expressing in the LR table is advantageous because the search can be performed quickly, and in particular, the LR table taking one look-ahead symbol is one of the syntax rules. This is advantageous because collation is easy. As the method of matching processing with the syntax rules performed by the grammar matching means, various matching methods can be adopted according to the format in which the syntax rules are compiled, but a shift / reduce parser such as an LR parser is used. The method to be performed is easy to realize, and in particular, the method to perform the extended LR parser that takes one look-ahead symbol is excellent in both high collation ability and easy realization.

[0014]

According to the speech recognition apparatus of the present invention, the speech input from the outside is recognized by the phoneme unit recognition unit in the unit of the phoneme unit, and the phoneme unit is recorded in the dictionary storage unit by the grammar matching unit. The syntactic rule compiled and recorded in the dictionary and the syntactic rule storage means is referred to and recognized by checking whether it is grammatically acceptable. Then, if necessary, an external editing instruction is given using the input means, and the dictionary managing means is caused to edit the contents of the dictionary in accordance with the editing instruction. Edits such as deleting registered words.

In this editing process, since the dictionary has a simpler structure than the syntax rules and is stored in the storage means independent of the syntax rules, the compilation process which takes a long time like the syntax rules is unnecessary. (Or it can be processed quickly even if compiled), and there is no need to do anything with regard to syntax rules. Therefore, the voice recognition can be restarted immediately after editing the contents of the dictionary. Therefore, in the voice recognition work, when a new word to be recognized occurs or a word that is no longer needed to be recognized thereafter, the contents of the dictionary can be immediately changed by the user's instruction. You can
The voice recognition can be performed efficiently.

Further, according to the speech recognition apparatus of claim 2,
While realizing the quick dictionary editing process as described above, the grammar collating means performs the efficient grammar collating process in which the LR parser predicts the classification of words such as parts of speech, and generally the highly efficient speech recognition process. Further, according to the speech recognition apparatus of claim 3, the grammar matching means predicts the word classification such as the word type by the LR parser, and further,
A succeeding phoneme piece is predicted from the information on the leading phoneme piece candidate and the phoneme piece having no word for each word category recorded in advance in the dictionary storage means. Then, the phoneme piece recognition means finds the predicted phoneme piece from the input voice data, and recognizes the phoneme piece of the subsequent word from the voice data. Even in this case, since the dictionary storage means is configured independently of the syntax rule storage means,
The rapid dictionary editing process as described above is also realized.

[0017]

DESCRIPTION OF THE PREFERRED EMBODIMENTS A voice recognition device according to a first embodiment of the present invention will be described with reference to the drawings. It should be noted that the present embodiment is a voice recognition device that implements a quick dictionary editing process and at the same time, performs grammatical matching on a sequence of recognized syllables of an input voice by predicting a part of speech of a word by an LR parser. The Japanese sentence obtained as is input to the word processor.

As shown in FIG. 1, the voice recognition device recognizes a microphone 1 for inputting voice from the outside and the input voice in syllable units and generates a syllable sequence (syllable string) as a candidate. Phoneme piece recognizing means 2, a grammar collating means 3 for collating whether or not the generated syllable string can be grammatically accepted by referring to a dictionary and syntax rules, a dictionary storing means 5 storing a dictionary 4, and a compiled form. Syntax storing means 7 storing the syntax rules (LR table) 6 and candidate selecting means 9 for narrowing down these candidates by referring to the concatenation evaluation table 8 when there are a plurality of grammatically matched candidates, and recognition. The word processor 10 to which the resulting sentence is input, the dictionary management means 11 for editing the contents of the dictionary 4 recorded in the dictionary storage means 5, and the word processor 1
0 and the input / output terminal 12 having an input / output device such as a keyboard and a mouse for the dictionary management means 11.

In the phoneme unit recognition means 2, a syllable HMM model 13 storing an HMM for each syllable, a signal processing means 14 for frequency-analyzing the electric signal of the voice data from the microphone 1 for each frame, and a frequency analysis. A feature amount extraction unit 15 for extracting a feature amount from voice data and a syllable collation unit 16 for collating the extracted feature amount with the HMM 13 to generate a candidate recognized for each syllable are provided. For example,
A syllable string as a recognition candidate as shown in FIG. 2 is generated by the phoneme unit recognition unit 2 and output to the grammar matching unit 3. It should be noted that the symbol $ in the figure indicates silence which means the end of the syllable string, and the number attached to each syllable indicates the score of syllable matching.

Here, in this embodiment, the above-mentioned dictionary 4,
LR table 6, linked evaluation table 8 and HMM model 13
Are separately stored in the storage device 17, and in particular, the dictionary 4 can be accessed by the dictionary management means 11 independently of the LR table 6 and the like.

The grammar collating means 3 holds the syllable strings sequentially generated by the phoneme unit recognizing means 2 as a candidate list as shown in FIG. 3, and the dictionary 4 and the LR table 6 for the syllable strings held in these candidate lists. Check the grammar while referring to. For each candidate in the candidate list, a syllable string in the dictionary that is not yet finalized as a word is stored as a “syllable string in dictionary dictionary”, which is a list of words already defined in the tree-structured syllable string shown in FIG. "Definite output notation" for saving the output notation, "prefetching sign" for saving the look-ahead symbol for classifying words (part of speech in this embodiment) necessary for parsing, and saving the LR parser parsing analysis state in the form of a stack The information of each item of "LR parsing state" and "score" which stores the score calculated up to the present time is stored.

The dictionary 4 recorded in the dictionary storage means 30
Includes a vocabulary accepted as an input to the word processor 10. As shown in a part of FIG. 4, a tree structure part (trie dictionary part) indicating a phoneme string and an output notation and word classification (part of speech) are used as words. It is composed of a table part described as information. That is, ○ in the tree structure part
The position of the syllable surrounded by a mark indicates that the syllable string up to the syllable can be accepted as a word, and the number attached to the position surrounded by a circle indicates an address to the table part. The address of the table portion means a reference such as a pointer or label for referring to the table, and does not always match the address of the storage device 17.

The syntax rule storage means 50 records a syntax LR table 6 composed of an action table and a Go-to table, a part of which is shown in FIG.
The LR table 6 is obtained by compiling the syntax rules described by the following context-free grammar expressed in a form that approximates the syntax rules of a sentence (especially Japanese, which is a natural language in this case) input to the word processor 10. Has become.

For example, the syntax rules to be compiled are as follows: Rule 0: sentence → bunsetsu (sentence is composed of clauses) Rule 1: bunsetsu → bunsetsu + bunsetsu (bunsetsu consists of a combination of clauses) Rule 2: bunsetsu → noun phrase ( Phrase consists of noun phrase) Rule 3: Noun phrase → noun + particle (noun phrase consists of combination of noun and particle) Rule 4: noun phrase → adjective + noun phrase (noun phrase is combination of adjective and noun phrase) Consists of) and so on.

In the action table shown in FIG. 5,
The leftmost column shows the state of syntax analysis, and the uppermost column shows the part of speech to be input. And when in a certain state,
When the part-of-speech of the next word is input as the look-ahead symbol, the column to be referred to in this action table is determined, and the action to be performed according to the instruction entered in each column is 4 types: acceptance, failure, shift, and reduce. Determined from within. Acceptance indicates that the syllable string input so far has been successfully matched as a sentence, and failure indicates that the syllable string input so far has not been recognized as a sentence. Note that in FIG. 5, failures are represented by blanks.

The shift indicates that the collation is in progress, and the operation of stacking a new state on the stack is performed. For example, shift 5 indicates stacking state 5 on. Reduce indicates that one rule can be applied, and the applicable rule among the syntax rules described above (for example, Rule 2 in Reduce 2).
The state is taken out from the stack by the number on the right side of, and the operation for stacking the next state is performed by referring to the Go-to table from the part of speech on the left side of the rule and the state of the top of the stack. Similar to the action table, the Go-to table is a table in which the heading in the leftmost column indicates the state and the heading in the uppermost column indicates the part of speech, and is referred to after the action table is reduced.

The candidate selecting means 9 is provided when there are a plurality of candidates as the voice recognition result output from the grammar matching means 3.
A process of narrowing down these candidates with reference to the connection evaluation table 8 is performed. The concatenation evaluation table 8 stores in advance the concatenation scores of two words, a part of which is shown in FIG. 6, as a criterion for the narrowing down process. For example, the score in which “Climb” is concatenated with “Asahi” is The possibility that two words are concatenated is ordered by score, such as −0.01, the score in which “shining” is concatenated with “Asahi” is −0.15, and the like. Therefore, the candidate selection means 9 selects from the character strings of the sentences and the like output as candidates from the grammar matching means 3 based on the maximum likelihood of word concatenation, and outputs the selected candidates to the word processor 10. .

The dictionary management means 11 edits the contents of the dictionary 4 stored in the dictionary storage means 5, and writes it in the dictionary 4 in accordance with the editing instruction input by the user using the keyboard of the input / output terminal 12 or the like. The information of the written word is searched for, and a new word is added or a word already written is deleted. The dictionary management means accesses only the dictionary storage means 5 (that is, the dictionary 4) and performs the editing process without affecting other stored contents such as the LR table 6 stored in the storage device 17.

The voice recognition apparatus of this embodiment will be described in more detail while explaining the operation of the voice recognition processing. First, the user utters a character string such as a sentence to be input to the word processor 10 by voice, and this voice is converted into an electric signal by the microphone 1 and input to the signal processing means 14. The signal processing means 11 converts the input audio signal into a digital signal that is easy to process, detects the start and end of the audio section by the power and duration of the audio signal, cuts out the audio signal, and cuts out the extracted section. Frequency analysis is performed for each unit time (frame) such as 10 ms.

In order to facilitate syllable matching,
The feature amount extraction unit 15 extracts the feature amount of the frequency-analyzed voice data and expresses it as a vector. The syllable matching is performed by the syllable matching means 16 by matching the above feature amount with the syllable HMM model 13, and the syllables that have been accepted with a score equal to or higher than the reference value are output in a tree structure as shown in FIG. In addition,
The syllable HMM model 13 is created in advance by the voice data for model creation and is stored in the storage device 17.

The syllable string candidates formed in this tree structure are subjected to dictionary lookup and syntactic analysis by the grammar matching means 3, and the character string candidates obtained as a result of the grammar matching are output to the candidate selecting means 9. Then, the optimum candidate selected by the candidate selecting means 9 is output to the word processor 10 and displayed on the display of the input / output terminal 12. That is, the most probable candidate is selected based on the score of the connection between words and the score at the time of syllable recognition described in the concatenation evaluation table 8, and the word string in which the candidate is confirmed is input to the word processor 10.

The grammar matching process by the grammar matching means 3 will be described in more detail with reference to the flowchart shown in FIG. First, a candidate list (FIG. 3) as a work area for dictionary lookup and syntax analysis is initialized (step S
1), it is determined whether the following processing has been performed up to the end of the branch for the tree-structured syllable string output from the syllable matching unit 16 (step S2). As a result, when the end is reached, the process is terminated, while when it is still in the middle, the tree-structured syllable string is sequentially read into the candidate list by tracing the branches (step S3). In this case, since the processing has started, all the first syllables in the tree-structured syllable string are read.

Next, by referring to the tree structure syllable string, it is judged whether or not a plurality of syllables follow the candidate syllable read in the candidate list (that is, whether or not the tree structure is branched) (step S4), and a plurality of syllables are obtained. If is followed by, the number of candidate syllables is copied by the number and stored in the candidate list, and the following syllable is stored at the end of the column of “syllable string in dictionary lookup” of the candidate list (step S5). If there are multiple initial syllables at the start of the process, prepare as many rows as the candidate list and store the first syllables in the "syllable string in dictionary lookup" column of each row. .

Next, a dictionary lookup process (described later with reference to FIG. 8) is performed on each candidate (each line) stored in the candidate list to determine the word (step S6).
As a result of the dictionary lookup process, the syntactic analysis process (described later with reference to FIG. 9) is performed only on the candidate for which the word is fixed (steps S7 and S8), and when the word is not fixed, it is stored in the candidate list. The dictionary lookup process is repeated for the next candidate (step S9). That is, the dictionary lookup process is performed for all candidates stored in the candidate list,
Parsing is performed on all the candidates whose words have been decided.

When the number of candidates stored in the candidate list is too large as a result of the above processing, the candidate having a low score has a low reliability as a recognition result, and hence the subsequent processing. In order to reduce the amount, candidates with a low score are deleted from the candidate list to narrow down the number of candidates (step S10). After that, the tree-structured syllable string that is the output result of the syllable recognition is referenced again, and the syllables that follow in the column of "syllable string in dictionary lookup" of the candidate list are further extended (step S3), and the above process is performed. This process is repeated until the tree structure syllable string is traced to the end or the candidate list becomes empty.

As a result of the above processing being completed,
If there are no candidates to be output, it means that the voice recognition has failed. In this case, it is preferable to display some message indicating that the recognition has failed on the display of the input / output terminal 12 to the user or give a voice notification by providing a voice synthesis function.

The above-mentioned dictionary lookup processing is performed according to the flowchart shown in FIG. 8 with reference to the dictionary 4 stored in the dictionary storage means 5. First, the tree structure portion of the dictionary 4 shown in FIG. 4 is traced (step S21), and it is determined whether or not there is a syllable string corresponding to the syllable string stored in the "syllable string being dictionary searched" column of the candidate list. (Step S22). As a result, if the tree structure portion cannot be traced, it is a candidate having a syllable string that is not in the dictionary 4, so that candidate is deleted from the candidate list and the dictionary lookup processing is ended (step S2).
3). On the other hand, when the tree structure portion is traced, it is determined whether the last syllable traced has a word acceptance label (circle mark) (step S24).

As a result, when the word acceptance label is not attached, it is necessary to continuously extend the subsequent syllables and perform dictionary lookup processing, so the candidate is saved as it is and the processing of the next candidate is performed. The dictionary lookup processing is ended. On the other hand, if the word acceptance label is attached, it means that the word has been fixed in the target syllable string stored in the "syllable string in dictionary lookup" column of the candidate list. After processing, the dictionary lookup processing is ended.

That is, the reference address to the table portion added to the word acceptance label in the dictionary 4 is traced, the table portion of the dictionary 4 is read (step S26), and the output notation is set in the final output notation column of the candidate list. It is added to the end (step S27), and the part of speech of the word is written in the look-ahead symbol column of the candidate list (step S28).

There are cases in which the candidates are divided into a plurality of candidates in the above-mentioned word determination process. There are a case where a plurality of words are fixed at the same time, and a case where a certain word is fixed and the dictionary lookup of another word continues at the same time. Both may occur at the same time. The former is a case where there are words having different parts of speech and output notations even with the same sound, and in this case, it can be detected by adding a plurality of reference addresses to the word acceptance label. The latter happens when another word is included as part of a long word. For example, when a syllable string “Asa” exists as a candidate in the dictionary 4 of FIG. 4, there is a candidate whose dictionary lookup is determined by “Asa” and a candidate which extends a branch to “Asahi”. When such a conflict occurs, the candidates in the candidate list are duplicated and treated as different candidates (step S29), and the above-mentioned word determination process (steps S26 to S28) is performed for each candidate. .

The above-mentioned syntax analysis processing is performed according to the flowchart shown in FIG. 9 with reference to the LR table 6 stored in the syntax rule storage means 7. In the syntactic analysis process, the extended LR parser syntactic analysis is performed on the candidate whose word has been determined based on the look-ahead symbol (part of speech) obtained by the dictionary lookup process and the stack of the state of the candidate. First, the action table of the LR table 6 is drawn based on the state of the top of the stack and the look-ahead symbol (step S3).
1). Here, in the extended LR parser parsing, one interpretation can be performed for a plurality of candidates, and in that case, there are a plurality of entries as a result of subtracting the action table. Therefore, it is judged whether there are a plurality of entries (step S32), and if there are a plurality of entries, the candidates are copied by the number of the entries (step S3).
3), proceed with the analysis according to each entry for each candidate.

Parsing is performed by the actions of reduce, shift, failure, and acceptance based on the result of pulling the action table, and the process according to the decision result of which of these actions is performed (steps S34 to S36). ). First, if the result of pulling the action table is reduce, the state stack is rewritten based on the syntax rule information to rewrite the state stack (step S37), and the Go-to table is pulled to stack the next state on the stack (step S37). S3
8) Draw the action table again (step S3
1), the process according to the result is repeated.

If the result obtained by subtracting the action table is a shift, the next state is stacked on the stack and the process of parsing the candidate is completed (step S3).
9). If the result obtained by subtracting the action table is blank, it means that the syntactic analysis has failed. Therefore, the candidate is deleted from the candidate list, and the work of the syntactic rule matching for the candidate is finished (step S40). If the result obtained by subtracting the action table is acceptable, the candidate is moved from the candidate list to the recognition success list for output to the candidate selecting means 9, and the syntax analysis process for the candidate is terminated (step S41). ).

Here, when the contents of the dictionary 4 stored in the dictionary storage means 5 are edited during the voice recognition process as described above or when the voice recognition process is not performed. Is processed according to the flowchart shown in FIG. First, when the user wants to register a new word not included in the dictionary 4, the user inputs a word registration instruction to the dictionary management means 11 from the keyboard of the input / output terminal 12 and the like, and the dictionary management means At 11, the word registration process is performed.

In the word registration process, as shown in the flowchart of FIG. 10, first, the syllable string, the output notation, and the part of speech of the word newly input by the user from the keyboard of the input / output terminal 12 are stored in the dictionary management means. 11 receives (step S51), searches the dictionary 4 based on the input syllable string, and searches whether a similar word such as a word having the same reading is already registered in the dictionary 4 (step S52). As a result, if already registered (step S53), the fact that the same word already exists in the dictionary 4 is displayed on the display or the like of the input / output terminal 12 (step S54), and the word registration process ends. .

On the other hand, although not exactly the same, similar words such as a word having a different part of speech in the same phoneme sequence or a word having a different output notation in the same syllable sequence (same synonym, different homophone) already exist in the dictionary 4. If so (step S53), the contents of the similar word (output notation, part of speech, etc.) are displayed on the display of the input / output terminal 12 (step S55), and the user confirms whether or not to perform the registration process. Take (step S5
6). On the other hand, when the same or similar words are not yet registered in the dictionary 4 (step S53), the user confirms whether or not the registration processing is performed to confirm the necessity of the registration processing ( Step S56).

As a result of the confirmation (step S57), if the user rejects the word registration, the word registration is terminated. If the user approves, the word registration consists of a part of the syllable string. It is determined whether the word is already registered in the dictionary 4 (step S58). As a result, when the syllable string of the word to be newly registered is not already registered in the dictionary 4,
A syllable is added to the existing syllable string in the tree structure portion of the dictionary 4 to extend the branch so as to form a syllable string of the word to be newly registered, and a word acceptance label is written on the last syllable (step S59).

Next, the area (address) of the front part of the dictionary 4 is secured, and the output notation and part of speech of the word to be registered are recorded (step S61), and the address to this front part is used as the word acceptance label of the tree structure part. (Step S6
1). By the above processing, a new word is registered in the dictionary 4 and immediately used for the dictionary lookup processing (step S6). Well, when registering a plurality of words, the above process is repeated.

When deleting a word already registered in the dictionary 4, the dictionary management means 11 performs the processing according to the flowchart shown in FIG. First, the user receives an instruction to delete a word input from the keyboard or the like of the input / output terminal 12 and information on the word to be deleted (output display, syllable string, information that can specify a word such as a part of speech) (step S71). Then, the dictionary 4 is searched based on the input information (step S72), and the corresponding word is found in the dictionary 4.
It is checked whether it is registered in (step S73).

As a result, if the corresponding word does not exist, a message to that effect is displayed on the display of the input / output terminal 12, and the word deleting process is terminated (step S74).
On the other hand, when the corresponding word exists, the syllable string, the part of speech, and the output notation of the word stored in the dictionary 4 are displayed on the display of the input / output terminal 12 (step S
75), the user confirms whether the deletion process should be executed (step S76). If there are a plurality of applicable words, all the words are displayed and the user is allowed to specify the words to be deleted, and only the specified words are deleted.

Then, when the deletion of the word is not approved by the user, the deletion processing is terminated, while when the deletion is approved, the processing of deleting the corresponding word from the dictionary 4 is performed as follows. That is, first, the address to the front part of the word to be deleted is deleted from the corresponding word acceptance label of the tree structure part of the dictionary 4 (step S77). Next, it is determined whether or not a word having the same syllable string as the word to be deleted among the words not to be deleted is registered in the dictionary 4 (step S78). As a result, if such a word does not exist, the word acceptance label corresponding to the word to be deleted is deleted (step S79), while the corresponding word column is deleted from the table portion of the dictionary 4 (step S8).
0) If such a word exists, the corresponding word column is deleted from the front part of the dictionary 4 without deleting the word acceptance label (step S80).

As described above, in this embodiment, the voice signal input by voice is recognized for each syllable and the result is recognized in the dictionary 4.
It can be recognized by performing syntax analysis while referring to, and the result can be input to the word processor 10. By using the syntax rule (the syntax LR table 6 expressed in the form of the LR table) that has been compiled in advance so as to improve the effect of the syntax analysis, the syntax analysis that generally requires the number of processing steps can be performed in a short time. It can be carried out.
Further, since the dictionary 4 is constructed independently of the syntax LR table 6, the dictionary 4 can be edited without changing the syntax LR table 6 by instructing the dictionary management means 11. It is possible to edit the dictionary 4 without performing a long-time process such as recompiling the rules. Therefore, in the voice recognition device that recognizes the voice of a complex sentence that requires syntactic analysis, it is possible to immediately perform dictionary editing such as word registration and deletion as needed even during the voice recognition process.

Next, a voice recognition apparatus according to the second embodiment of the present invention will be described with reference to the drawings. In addition, the above-mentioned first
The same functional means as those in the embodiment of FIG. In the present embodiment, while automatically realizing a quick dictionary editing process, the phoneme sequence of the input speech is predicted for each word type (word classification) to realize quick and reliable phoneme unit recognition. It is a voice recognition device, and inputs a Japanese character string obtained as a recognition result to a database search system as a command sentence for database search.

The speech recognition apparatus of this embodiment has a structure as shown in FIG. 12, and the phoneme piece recognition means 2 has a phoneme collation means corresponding to the recognition of input speech in phoneme units. 26 and a phoneme model 23 are provided. Also,
The grammar matching unit 3 includes a matching control unit 31, a dictionary matching unit 32, and a syntax rule matching unit 33. The grammar matching unit 3 receives the phonemes recognized by the phoneme unit recognition unit 2 and stores the dictionary storage unit 5 and the syntax rule memory. By referring to the contents of the means 7, grammatically succeeding phonemes are predicted, and grammatical matching is performed in a format instructing the phoneme piece recognition means 2 whether the predicted phoneme can be recognized next.

The dictionary storage means 5 stores information on the vocabulary included in the statement to the database search system 41 as shown in FIG.
A tree structure part showing a phoneme string of a word, a table part storing output display and type of the word, and a head phoneme that can be the head of the phoneme string corresponding to each word type are stored. It is stored as a head phoneme table classified by part of speech. Since the present embodiment accepts a command sequence for searching the database 42 instead of the natural language, unlike the first embodiment, the word classification in the dictionary 5 and the syntax rules 7 is linguistic. It is not the part of speech that is commonly called, but the type of word. For example, the names of commands that take arguments, the names of commands that do not take arguments, and the types of item names and titles of data on the database are used as word categories.

As in the first embodiment, the syntax rule storage means 7 describes the syntax rules of the command string (that is, the imperative sentence) for the database search system 41 by the context-free grammar and compiles it into the syntax LR table. And stored. The database search system 41 is a system for searching the database 42 according to a command string, and operates by the input from the acquisition terminal 12 and also by the voice recognition result (command string) output by the candidate selecting means 9. In the present embodiment, the dictionary storage unit 5, the syntax rule storage unit 7, and the phoneme model 23 are each configured as an independent storage device, and the dictionary storage unit 5 is independent of the syntax rule storage unit 7. It can be accessed by the management means 11.

The voice recognition apparatus of this embodiment will be described in more detail while explaining the operation of the voice recognition processing. When the user utters a command string instructing the database search system 41 by voice, this voice is converted into an electric signal by the microphone 1 and input to the signal processing means 14. The signal processing unit 14 converts the input audio signal into a digital signal and then frequency-analyzes each frame, and the feature extraction unit 15 extracts a feature amount from the frequency-analyzed voice data.

The phoneme collating means 26 activates a necessary phoneme model of the phoneme model 23 for the phoneme predicted by the grammar collating means 3 as described later, and outputs a voice signal in which a trace amount is extracted. The list of phonemes that have been matched and accepted with a score equal to or higher than the reference value is the matching control means 3 of the grammar matching means 3.
Output to 1. Then, the grammar collating means 3 collates the phoneme string recognized by the phoneme piece recognizing means 2 with the dictionary stored in the dictionary storage means 5 or the LR table stored in the syntax rule storage means 7. Here, in the present embodiment, first, the phonemes (groups) that can be accepted next are predicted by the LR parser syntax analysis, and these predicted phoneme groups are passed to the phoneme matching means 26, and as a result, the phoneme matching means 26 is The processing proceeds by a method in which the grammar matching means 3 receives the received phonemes.

The collation control means 31 has a candidate list, an example of which is shown in FIG. 14, as a work area. The candidate list is a "number" of the label of the candidate, a "fixed output notation" that records the output notation of the already fixed word (command),
"Phoneme string in dictionary lookup" that records a phoneme string in which a word has not yet been determined, "LR parsing state" that records the state of LR parser parsing as a stack, and the type of a confirmed word in LR parser parsing Each column includes a “look-ahead symbol” recorded as a look-ahead symbol, a “predicted phoneme” recording a predicted phoneme to be matched next, and a “score” recording a cumulative score of the candidate.

The collation processing by the grammar collating means 3 will be described in more detail with reference to the flow charts shown in FIGS. First, the candidate list is initialized, and referring to the part-of-speech head phoneme table of the dictionary shown in FIG. 13, the phoneme starting from the kind of word that may come at the head of the command string is set as “predicted phoneme” of the candidate list. Is entered in the field of "" (step S101). For example, if the type of word that may come at the beginning of the command string is an item name, then / d /, / k /, / j /, and / m / are each "predicted phoneme" from the beginning phoneme table by part of speech. The candidates entered in the column are created in the candidate list.

Then, the following process is repeated until the candidate list becomes empty (step S102). First, the phoneme matching unit 26 is requested to start matching of the predicted phonemes of each predicted candidate (step S103). The phoneme matching unit 26 performs a process as to whether or not the predicted phoneme can be matched at the head of the input speech, and the grammar matching unit 3 accepts the result list of the phonemes that have been successfully matched or have failed (Step S104). ). Next, the candidate in which the matching failed phoneme is entered in the “predicted phoneme” column is deleted from the candidate list (step S105), and the successfully matched phoneme is selected as the candidate entered in the “predicted phoneme” column. In contrast, Figure 1
A process of predicting the next phoneme that can come next, which will be described later with reference to FIG. 6, is performed (step S106).

The above process is repeated until the candidate list becomes empty. As a result, if there is no candidate in the recognition candidate list, the recognition has failed. If there is a candidate, the candidate selecting means 9 gives the highest score. Is selected as the recognition result (step S107), and the command string obtained as the recognition result is input to the database search system 41 (step S108).

The process of predicting the next phoneme (step S106) will be described with reference to the flowchart shown in FIG. First, since the previously predicted phoneme has been successfully confirmed and confirmed, the previous phoneme is moved from the "predicted phoneme" field of the candidate list to the end of the "phoneme string in dictionary lookup" field. (Step S111). Here, as in the first embodiment, the word may be divided into a word that has not been fixed and a word that has not been fixed. Even when a word is fixed, a plurality of words may be fixed in the same manner. According to the number of all such cases, the candidate is copied to be another candidate (steps S112 and S113). Then, for each of the copied candidates, the unprocessed candidates are taken out one by one and the following processes are sequentially performed (step S114).

When the word is not yet determined for the unprocessed candidate (step S115), the phoneme connected to this is listed up as the next predicted phoneme by referring to the tree structure part of the dictionary (step S115). S116). On the other hand, when a word is determined, a phoneme prediction process including a syntax analysis described below is performed based on FIG. 17, and a phoneme that may possibly be at the beginning of the next word is listed as a predicted phoneme ( Step S117).

Then, as many candidates as the number of listed predicted phonemes are copied to the candidate list (step S118),
Predicted phonemes are entered one by one in the "predicted phoneme" column of the candidate list of each candidate (step S119).
The above processing is performed until there are no unprocessed candidates in the candidate list (step S120), and based on these predicted phonemes, the phoneme that follows the phoneme sequence or the first phoneme of the word that follows in the word sequence is phoneme-matched. The unit 26 collates (step S104).

The phoneme prediction process (step S117) including the above-mentioned syntax analysis will be described with reference to the flowchart shown in FIG. First, since the word has been determined (step S115), the word data is read from the front part of the dictionary by referring to the address added to the word acceptance label of the tree structure part of the dictionary (step S131). And
The output notation of the extracted word is added to the end of the "fixed output notation" column of the candidate list (step S132), and the type of the extracted word is also added to the "look-ahead symbol" of the candidate list.
Is filled in in the column (step S133). Further, the phoneme string in the "phoneme string in dictionary lookup" column of the candidate list is cleared (step S134).

Then, using the type of word (corresponding to the part of speech in the first embodiment) entered in the "look-ahead symbol" column, the same syntax analysis processing as in the first embodiment shown in the flowchart in FIG. 9 is performed. Is performed (step S135). As a result of this syntactic analysis, the result of the accepted or failed candidate is fixed, so that no further processing is required (step S13).
6) In other cases, the action table is referenced to list the types of words that come next from the current LR parsing state that do not fail (step S13).
7). Then, with respect to the listed word type, the head phoneme table for each part of speech in the dictionary is referred to, and the next predicted phoneme is listed (step S138). That is, a phoneme that may possibly come at the beginning of this word is selected as a predicted phoneme from the types of the following words, and the phoneme matching unit 26 is caused to perform matching processing based on this predicted phoneme (step S1).
04).

As described above, a command string is input by voice to the database search system 41. During this work or work suspension, new data is registered in the database 42 and item names etc. are newly added. Then, it becomes necessary to add the recognizable word to the dictionary stored in the dictionary storage means 5. This word registration processing is automatically performed by the dictionary management means 11 according to the flowchart shown in FIG.

For example, when a new item name is added to the database 42, the item name to be added and its type are sent from the database search system 41 to the dictionary management means 11. The dictionary management means 11 is an input / output terminal 12
Then, the user is inquired about the reading method, and the reading method corresponding to the item name is acquired (step S140). Next, it is checked whether or not the same item name (word) in the dictionary is already registered (steps S141 and S142), and if the same item name is already registered, the process is terminated.

On the other hand, if there is no same item name, the first
A new item name is registered in the dictionary in the same manner as in the above embodiment. That is, a new address is acquired in the front part of the dictionary, and information on the item name to be registered is registered (step S
143), if necessary, the branch is extended to the tree structure portion of the dictionary, and the address of the front portion is added to the word acceptance label (step S144). Then, in this embodiment, the head phoneme table for each part of speech in the dictionary is checked, and when the head phoneme of the item name registered in this table is not included, the phoneme is additionally written (step S145). As a result, a new word can be registered in the dictionary stored in the dictionary storage means 5 without any influence on the syntax rule storage means 7.

As described above, in the present embodiment, it is possible to recognize a voice signal input as a voice by performing syntax analysis while referring to a dictionary and input the result as a command string to the database search system 41. it can. In particular,
In this embodiment, in addition to the features of the first embodiment, even when the dictionary storage means 5 and the syntax rule storage means 7 are separated,
The phoneme to be matched can be predicted. Then, by predicting the phonemes, the number of phoneme models activated for the matching process can be reduced, and the processing amount required for the matching can be reduced. In addition, since the database search system 41 automatically registers words via the dictionary management means 11 with the update of the contents of the database 42, the vocabulary that can be accepted is automatically increased according to the change of the contents of the database 42. be able to.

[0072]

As described in detail above, according to the speech recognition apparatus of the present invention, the dictionary management is performed so that the dictionary can be made independent from the syntax rules that require a long time to compile and the dictionary can be accessed without affecting the syntax rules. With the provision of means, it is possible to immediately make changes such as addition or deletion of new vocabulary to the dictionary by external instructions without inputting standard voices and reconstructing syntax rules as needed. ,
It is possible to realize efficient voice recognition.

Further, according to the voice recognition device of claim 2,
To achieve efficient speech recognition as a whole by performing efficient grammar matching processing that predicts the classification of words such as parts of speech by the LR parser in the grammar matching means while realizing the above-described quick dictionary editing processing. You can Still further, according to the speech recognition apparatus of claim 3, the LR parser predicts a subsequent phoneme piece, and the phoneme piece recognition unit recognizes the phoneme piece from the input voice data based on the predicted phoneme piece. As a result, efficient phoneme matching is realized,
It is possible to realize highly efficient voice recognition as a whole.

[Brief description of drawings]

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

FIG. 2 is a conceptual diagram showing an example of a phoneme string obtained by phoneme piece matching.

FIG. 3 is a conceptual diagram showing an example of a candidate list.

FIG. 4 is a conceptual diagram showing the contents of a dictionary.

FIG. 5 is a conceptual diagram showing contents of an LR table in which syntax rules are compiled.

FIG. 6 is a conceptual diagram showing the contents of a connection evaluation table.

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

FIG. 8 is a flowchart showing a procedure of dictionary lookup processing according to the first embodiment of the present invention.

FIG. 9 is a flowchart showing a procedure of syntax analysis processing according to the first example of the present invention.

FIG. 10 is a flowchart showing a procedure of word registration processing according to the first embodiment of the present invention.

FIG. 11 is a flowchart showing a procedure of word deletion processing according to the first embodiment of the present invention.

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

FIG. 13 is a conceptual diagram showing contents of a dictionary according to a second embodiment of the present invention.

FIG. 14 is a conceptual diagram showing an example of a candidate list according to a second example of the present invention.

FIG. 15 is a flowchart showing a procedure of voice recognition processing according to the second embodiment of the present invention.

FIG. 16 is a flowchart showing a procedure of phoneme prediction processing according to the second embodiment of the present invention.

FIG. 17 is a flowchart showing a procedure of a syntax analysis process according to the second embodiment of the present invention.

FIG. 18 is a flowchart showing a procedure of word registration processing according to the second embodiment of the present invention.

[Explanation of symbols]

2 ... Phoneme piece recognition means, 3 ... Grammar matching means,
5 ... dictionary storage means, 7 ... syntax rule storage means,
11 ... Dictionary management means, 12 ... Input / output terminal,

Claims (3)

[Claims] 1. A syntax rule storage means for storing a syntax rule for defining an acceptable speech grammar in a compiled form, and a syntax rule storage means configured independently of each other, wherein an acceptable word is at least phoneme string notation. , The output notation, the dictionary storage means for recording the classification of words, the phoneme unit recognition means for recognizing the input speech in units of phoneme units, and the recorded contents of the dictionary storage means and the syntax rule storage means, Grammar collating means for collating the arrangement of the pieces in terms of grammar, inputting means for inputting a word editing instruction to the dictionary storing means from the outside, and recorded contents of the dictionary storing means based on the instruction from the inputting means A voice recognition device comprising: a dictionary management unit for changing the. 2. The grammar checking means refers to the recorded content of the dictionary storage means for the recognized phoneme string sequence to determine the corresponding word, and refers to the recorded content of the syntax rule storage means for the word. Then, grammatical matching is performed by LR syntax analysis, and the classification of the subsequent word is predicted based on the classification of the word, and the process of determining the subsequent word using the predicted classification of the word is performed. The voice recognition device according to claim 1, wherein the voice recognition device is a voice recognition device. 3. A dictionary storage means stores in advance a candidate of a phoneme piece which becomes a head for each classification of words, and the grammar matching means records the content of the dictionary storage means for the recognized phoneme piece string. To determine the corresponding word, refer to the recorded content of the syntax rule storage means for the word, perform grammatical matching by LR parsing, and classify subsequent words based on the classification of the word. And the subsequent phoneme piece from the recorded content of the dictionary storage means based on the classification of the word, and the phoneme piece recognition means finds the input phoneme by finding the predicted phoneme piece at the beginning. The voice recognition device according to claim 1, wherein the voice recognition device recognizes the voice.