JPH03180899A - Voice recognizing method - Google Patents

Abstract

PURPOSE:To analyze a sentence efficiently for sentence voice recognition by providing a clause sequence storage part, a clause detection start condition generation part, etc., and carrying out processes for clause detection and meaning analysis simultaneously, progressively, and integrally. CONSTITUTION:The clause detection part 12 of a meaning expression output part 11 detects a clause candidate from the phoneme sequence of an input sentence voice recognized by a phoneme recognition part 10 and a meaning analysis part 13 checks the possibility of linguistic and position-concerned connections between the clause candidate and a clause sequence group which is already stored in a clause sequence storage part 14, so that clauses which can be connected are stored in the storage part 14 together with the meaning expression of modification analytic results. At the same time, the clause detection start condition generation part 15 generates clause start conditions according to the storage contents of the storage part 14 and supplies them to a clause detection part 12, and clause candidates which do not match the conditions are inhibited from being detected. Those clause detecting process and meaning analyzing process are carried out simultaneously, progressively, and integrally to inhibit unnecessary clause detection from being performed, thereby analyzing sentences efficiently.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition method when the main object of recognition is continuously uttered sentence speech.

BACKGROUND OF THE INVENTION Conventionally, various methods have been proposed for recognizing sentence speech.

One typical method when recognizing Japanese sentences is to first estimate which subsections of the input speech correspond to which clauses for all speech intervals, and then create a set of clause candidates. The set of phrase candidates consists of a series of phrases that satisfy the temporal positional relationship and linguistic constraints such as grammar and meaning (in general, the overlap between phrases often refers to rules). This is a method of estimating a subset as a recognition result.

A set of clause candidates with temporal position information is
If we call it the "Bunsetsu lattice," there are generally multiple sets of clauses that simultaneously satisfy the above constraints obtained from the "Bunsetsu lattice." It is common practice to prepare in advance a scale (described later) derived from the above, rank the phrase candidates according to the evaluation value, and select the first-ranked candidate as the recognition result.

By the way, in order to realize the method described above, it is necessary to clarify the following specific issues.

(1) A specific method to obtain the final result from the bunsetsu lattice.

(2) A method of creating a bunsetsu lattice from audio signals.

(3) Type of grammar used.

As for (1), which is the subject of the present invention, conventionally, “
This method uses a brute force method to check the temporal positional relationships and linguistic constraints of all the clauses that make up the clause lattice, and a tree search or path search method in the field of graph theory from the beginning of the sentence to the end. , a method is known in which the above-mentioned linguistic constraint check is applied to clause candidates from the beginning to the end of the sentence (for example, "Speech Recognition" by Niimi, published by Kyoritsu Shuppan, published in 1979).

However, the former is due to computational efficiency, and the latter is due to the linguistic structure of Japanese, in which the clause corresponding to ``kari'' in ``modariuke'' is placed before the clause corresponding to ``uke'', a language unique to Japanese. Each method has problems in terms of consistency with respect to its structure, and it has not always been possible to obtain satisfactory results when realizing Japanese sentence speech recognition.

In order to solve this problem, the present inventor proposed a method to obtain a sequence of clause candidates by searching from the end of the sentence to the beginning of the sentence, subjecting the dependencies between clauses to the constraints of the structure rules. Proposed.

In other words, this method has the advantage of being suitable for the language structure of Japanese and being efficient.

That is, FIG. 3 shows a sentence speech recognition system using this method,
The phoneme recognition unit 1 has means for performing phoneme recognition from an input audio signal and outputting a phoneme sequence. The obtained phoneme sequence is not guaranteed to be completely correct, and has a certain probability of containing errors due to the speaker, ambient noise, or the performance of the phoneme recognition device itself.

Therefore, it is not possible to directly apply conventional syntactic analysis and semantic analysis techniques based on natural language processing technology for character input, and a processing method specific to sentence speech recognition is required.

The clause lattice creation unit 2 has a function of obtaining clause candidates by estimating which part of the recognized phoneme sequence containing errors is likely to correspond to which clause. The estimated phrase is accompanied by temporal position information regarding the beginning and end, as well as information regarding the validity of the estimation. The validity of phrase estimation is given, for example, by the likelihood derived from a concussion matrix between phonemes obtained from each error rate of phoneme substitution, addition, and omission. Here, it is assumed that all phrase candidates whose estimated validity exceeds a predetermined value over the entire range from the beginning of the sentence to the end of the sentence are obtained.

The semantic analysis unit 3 obtains a series of clauses as the final result by concatenating clause candidates in the "clause lattice" according to positional relationship constraints and linguistic constraints. The grammar (rules) that determine linguistic constraints are:
For example, Fillmore's case grammar (described later), 5c
A concept-dependent grammar using hank can be considered.

FIG. 4 shows an example of a phoneme recognition result and an example of a ``bunsetsu lattice'' when an example sentence ``Please give me three tickets to Shin-Osaka'' is input by continuous utterance. In the figure, the phrases with a/dare lines in the "Phrase lattice" are the correct phrases.

Also, in this example, a case grammar is used to check linguistic constraints, so a brief explanation of the case grammar will be given below.

Case grammar is a grammar that describes how predicates and other words coexist. In other words, it is a method of analyzing sentences that does not consider the roles of subjects, predicates, objects, and complements in a sentence, but rather considers the position of each word in terms of meaning for the predicate. As mentioned above, case grammar emphasizes semantic analysis rather than syntactic analysis, so when considering its application to (Japanese) sentence speech recognition,
It has the following advantages.

(1) It is easy to narrow down word candidates using semantic co-occurrence relationships between words.

(2) It is easy to handle Japanese because there is a large degree of freedom regarding word order.

(3) Since the results are obtained in the form of meaning, it is easy to incorporate into dialogue systems, etc.

FIG. 6 shows an example of the result of semantic analysis using this case grammar, where (MODE MEIIIEI) in the -th line indicates that this sentence is an imperative sentence. Also, the second line (
7) (ACT KIJDASARU) indicates that the predicate means the act of ``giving.'' Furthermore, the third line (OBJECT KIPPU) and the following are descriptions of phrases or clauses that have the function of explaining the details of the action of the predicate "Gaseru". That is, the act ``
The object of ``Give me'' is ``tickets'', the number of which is ``3'', and the destination is ``Shin-Osaka''.

In case grammar, for each verb, items are prepared that can explain the meaning in detail, and semantic analysis proceeds by sequentially applying each word in a sentence to the corresponding item. These items that exist for each verb are called ``case slots,'' meaning ``grooves for filling.''
It is called. Also, the operation of sequentially applying each word in a sentence to the corresponding item is called ``slot filling'', which means ``slot filling operation.'' These terms will also be used hereinafter.

In the embodiment of the present invention described later, the above-mentioned linguistic restriction check refers to a check as to whether "slot filling" is accepted.

Note that the language processing system handled by the present invention not only handles case relations between clauses as explained above, but also adjunctive modification relations such as the example of "red flower", conjunctive modification relations such as "blooms beautifully", Also, conjunctions such as ``buy a ticket and go to Kyoto'' are treated with the same concept as ``slot filling'' in case grammar. In other words, in the present invention, not only conjugated words but also modified words have slots, and modification is a framework for analyzing relationships.

The process of concatenating clause candidates in the "Bunsetsu lattice" is performed in stages from a predetermined number of clause candidates located at the end of the sentence toward the beginning of the sentence.

Problems to be Solved by the Invention However, in the case of the above-mentioned conventional method, it is difficult to
At the stage of creating a bunsetsu lattice, clause candidates that are unnecessary for semantic analysis are also detected, which poses the problem of reducing the efficiency of recognition processing.

The present invention has been made in view of these conventional problems, and its purpose is to integrate the process of creating a "bunsetsu lattice" and the process of semantic analysis, which were conventionally performed as separate processes, To obtain a speech recognition method that avoids unnecessary detection of phrase candidates and realizes an efficient segmentation analysis mechanism by providing a new means for simultaneously performing phrase detection and semantic analysis processing. There is a particular thing.

Means for Solving the Problems In order to achieve the above object, the speech recognition method of the present invention includes a phrase detection means for detecting section information and reliability information of phrase candidates from an audio signal, and a phrase detection means for detecting interval information and reliability information of phrase candidates from a speech signal. Created by connecting phrase candidates as a series of phrases that satisfy the constraint conditions between phrases with a semantic analysis means that examines constraints on the positional relationship between phrase candidates in the candidates and linguistic constraints such as grammar or meaning. and a clause series storage means for storing a clause series to be analyzed, and based on the information of the end point of the clause series that is being analyzed, which is stored in the clause series storage means,
It is proposed that the sentence analysis process be carried out by detecting clause candidates in only the necessary sections using the clause detecting means.

Effects The above-mentioned speech recognition method realizes efficient recognition of Japanese sentence speech with higher recognition accuracy according to the following effects.

(1) By selecting clause candidates in stages, it is possible to perform analysis processing while simultaneously creating a plurality of clause sequences.

(2) At the processing stage of (1), “dependency” between bunsetsu clauses
Analysis can be completed.

(3) For phrase sequences whose estimated validity value does not reach a certain evaluation standard during the course of analysis, subsequent processing can be discontinued.

(4) The analysis process proceeds from the end of the sentence to the beginning of the sentence, and the search can be performed under control based on clause estimation validity and linguistic constraints.

EXAMPLE Hereinafter, an example of the present invention will be described in detail with reference to FIG.

FIG. 1 shows the concept of a sentence speech recognition system according to the present invention. The phoneme recognition unit 10 is a means for performing phoneme recognition from an input speech signal and outputting a phoneme sequence. The phoneme sequence has a certain probability of being incorrect due to misrecognition. Contains.

The semantic expression output unit 11 indicated by the dotted line in the figure is composed of a phrase detection unit 12, a meaning analysis unit 13, a phrase sequence storage unit 14, and a phrase detection start condition creation unit 16, and is a semantic expression that is a recognition result from a phoneme sequence containing errors. It has the function of obtaining Then, each component of the semantic expression output unit 11 avoids the detection of clause candidates that are unnecessary for semantic analysis, which occurs when calculating the conventional "clause lattice", and realizes efficient sentence analysis processing. It is supposed to be done.

Next, the operation of the sentence speech recognition system having the same configuration will be explained. In the phrase detection unit 12, the phrase series storage unit 1, which will be described later,
By referring to the information on the detection start position and the phoneme information in the vicinity of the detection start point in the information on the bunsetsu sequence start conditions created by the bunsetsu detection start condition creation unit 16 and the bunsetsu detection start condition creation unit 16, Detect candidates. Next, in the semantic analysis section 13, it is possible to link the obtained clause candidates linguistically and positionally with the elements of the clause series group stored in the clause series storage section 14 and accumulated up to the time of the current analysis. If connection is possible, the clause candidate is connected to the element of the clause series group and stored in the clause series storage unit 14 as a new clause series. The same rainbow as this, "Kariuke"
The semantic expression that is the analysis result is also stored in the phrase series storage unit 14. In addition, the phrase detection start condition creation unit 16 includes information on the phrase detection start position required by the phrase detection unit 12,
Information on phonemes near the phrase detection position is stored in the phrase sequence storage unit 14.
It also has the ability to create.

The above operations are started from a clause candidate having an end point at either the start or end position of the input speech, and are continued until the analysis of clause candidates having an end point at the other end position is completed. When the processing is completed, the phrase series with the highest validity evaluation value of the phrase series is set as the recognition result,
At the same time, the semantic expression of the form already shown in FIG. 4 can be obtained as a result of semantic analysis.

FIG. 2 shows an example of an algorithm that implements the sentence recognition function realized with the configuration shown in FIG. The example shown in Figure 2 explains the case where the analysis process proceeds from the end of the sentence to the beginning of the sentence. Of course, it is also possible to proceed with the analysis process in this direction. In preparation for the explanation, terms and symbols used in FIG. 2 will be defined.

Definition of Terms; [Definition 1] "The clause candidate satisfies the sentence-final condition" means that the clause can exist at the end of the sentence both grammatically and positionally.

[Definition 2] "The phrase candidate satisfies the sentence-initial condition" means that the phrase can exist at the beginning of a sentence both grammatically and positionally.

[Definition 3] “The bunsetsu candidate can be left connected to bunsetsu series B” means:
This means that it is possible to grammatically connect the bunsetsu s to the left end of the bunsetsu series B.

Definition of symbols; (Bn): Set of clause sequences with n clauses (F): Set of sentence candidates (b) Set of two clause candidates []: bleil that gives priority to operations in []: Sentence-final condition [ Clause candidate beam (B, ) that satisfies [Definition 1]: (B, l: An element of (B, ) which is a set of clause series that were not (remained) subject to pruning (B, l: An element of (B, ) Sentence-initial condition [Set of clause sequences that satisfy definition 2 adjIe(t (B, ): at least (B,
A set of phrase candidates that can be left connected to one element of A set of phrase sequence candidates created by concatenating sentence sequence and sentence sequence lJ: Union set operation In the algorithm shown in Figure 2, the number of phrases is n-1 (n
One unit is the process of connecting one clause to the clause series of ≧1) to create a clause series of N clauses. This one
Processing for one unit is expressed as "n-th clause processing" in the figure, and this process is repeated from n = 1 until the end condition shown in the figure is satisfied.

Next, to explain the relationship between the embodiment of the present invention shown in FIG. 1 and the algorithm according to the present invention shown in FIG. detection"
It is carried out by Detection of a set of clauses [bl is the clause candidate b jail at the end of the sentence in Figure 2. The function of the symbol adjleft, which expresses the adjacency condition between the clause sequence and the clause candidate, is as follows.
This is realized by the clause detection start conditions and the "semantic analysis" output in "Create Clause Detection Start Conditions" in Figure i11. In other words, the positional adjacency relationship between the bunsetsu sequence and the bunsetsu candidate is determined by the ``bunsetsu detection start condition'', and the semantic adjacency relationship is determined by the ``semantic analysis'' as a condition regarding the boundary between the bunsetsu sequence and the bunsetsu candidate. Ru. Furthermore, the phrase sequence at the time immediately after adopting the n-th phrase represented by the symbol 11n in FIG. 2 is stored in the "Phrase series storage" in FIG. It is used as a phrase candidate selection condition for the process of selecting one phrase.

Although the embodiments of the present invention have been described above, the present invention is not necessarily limited thereto. That is, for example, speech that involves the omission of words or phrases that occur in dialogue, or sentence speech that is uttered divided into phrases are also included in the recognition targets of the present invention.

Effects of the Invention As described above, according to the present invention, the process of creating a "bunsetsu lattice" and the process of semantic analysis, which were conventionally performed as separate processes, are integrated, and a new process of detecting and analyzing the clauses is performed. By performing this simultaneously with the processing, it is possible to avoid unnecessary detection of phrase candidates in conventional methods and realize an efficient sentence analysis mechanism.

[Brief explanation of drawings]

FIG. 1 is a functional block diagram of a sentence speech recognition system according to an embodiment of the present invention, FIG. 2 is a diagram showing an example of an algorithm when the present invention is applied to a sentence speech recognition system, and FIG. 3 is a block diagram of a conventional sentence speech recognition system. A functional block diagram showing a typical configuration example of a sentence speech recognition system. Figure 4 is a diagram showing the phoneme recognition results for sentence sounds and an example of a "bunsetsu lattice." Figure 6 is a diagram showing the semantic analysis results. It is. 12... Clause detection section, 13... Semantic analysis section, 14.
...Bunsetsu series storage section.

Claims (1)

[Claims] A clause detection means detects interval information and reliability information of clause candidates from the audio signal, and constraints regarding the positional relationship between clause candidates among the clause candidates obtained by the clause detection means and linguistic information such as grammar or meaning. The system includes a semantic analysis means for examining constraint conditions, and a clause series storage means for storing a clause series created by connecting clause candidates as a series of clauses satisfying the constraint conditions between clauses. A speech recognition method characterized in that the sentence analysis process is proceeded by detecting phrase candidates in only necessary sections by the phrase detecting means based on information on the end points of a phrase series that is being analyzed.