JPH077273B2 - Syntax processor for continuous speech recognition - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of use] The present invention receives an input word candidate from an acoustic processing device that acoustically processes an input speech by a continuous speech recognition device and applies a syntax rule to obtain a speech recognition result. Regarding a syntax processing device.

[Conventional technology]

Figure 7 shows, for example, literature (Kyoritsu Shuppan, Information Science Course E.19.
FIG. 3 is a functional block diagram of a conventional continuous speech recognition syntax processing device shown in 3, Speech Recognition, pages 139 to 142, 1979).

In the figure, (20) is a syntax rule storage means (60) for a word string candidate (22) obtained by reading an input word candidate (11) obtained by acoustically processing an input voice from a word string candidate storage means (50). A new word string candidate (22) is formed by connecting the input word candidate (11) to the word string candidate (22) that satisfies the connection condition by applying the syntax rule (33) searched from A syntax determination means for additionally storing in the string candidate storage means (50), and (40) is a word string having a minimum total acoustic distance value from the word string candidate storage means (50) when all input word candidates (11) have been input. A recognition result judging means for reading out the candidate (22) and outputting it as a voice confirmation result (77), (50) a word string candidate storing means for storing the word string candidate (22) generated by the syntax processing,
Reference numeral (60) is a syntax rule storage means for storing the syntax rule (33) described in the syntax state transition network.

In the syntax processing device for continuous speech recognition of the above-mentioned conventional example, first, the syntax determination means (20) processes forward speech (left-to-ri) from the beginning of the input speech along the time axis direction as shown in FIG.
The word name w of the input word candidate (11) obtained by acoustically processing the input speech by the ght method) is accepted in the final state p of the word string candidate (22) read from the word string candidate storage means (50). The syntax rule (33) retrieved from the syntax rule storage means (60) is applied to perform a syntax check. As shown in FIG. 9, for all of the word string candidates (sentence hypotheses) that are accepted / generated at this time, the one having a large total acoustic distance value (total evaluation value of the acoustic distance value) of each of the constituent words, that is, the likelihood is low. A beam search method (beam seam) that rejects objects and adopts only the smaller ones, that is, the ones with higher likelihood as final word string candidates at that time and processes the next input word candidates (11).
rch method) for hypothesis testing. The numbers attached to the nodes in FIG. 9 indicate the likelihood, the circles indicate the hypotheses that advance the test, and the black dots indicate the hypotheses that are rejected. As a result of this syntax test, a new word string candidate (22) formed by connecting the input word candidate (11) to the word string candidate (22) that satisfies the connection condition is additionally stored in the word string candidate storage means (50). To do.

Next, when the utterance ends and the input word candidates (11) are all input, the recognition result determination means (40) causes the word string candidate storage means (5).
The word string candidate (22) having the smallest total acoustic distance value is read from 0) and output as the voice recognition result (77).

For the method using the forward method as the processing direction, the most probable word candidate is stored as the initial hypothesis from among all the input word candidates (11) that were initially stored, as shown in Fig. 10. There is also a method of using an island-driven method in which the remaining part is processed in the front-back direction of the time axis while expanding from the beginning of the sentence to the end of the sentence.

[Problems to be Solved by the Invention]

In the conventional continuous speech recognition syntax processing device as described above,
Since the method using the forward method processes the evaluation and verification of the hypothesis in the time axis direction, a correct recognition result cannot be obtained if the correct hypothesis is once rejected in the middle, and the word candidates near the beginning of the occurrence have acoustic distance values. Is not necessarily small, it is necessary to make many hypotheses in order to obtain correct recognition results. Furthermore, in order to reduce the rejection in the sentence and improve the sentence recognition rate, words with a relatively large acoustic distance value are used. It is necessary to handle the candidates in the same way, and the processing amount becomes huge.

On the other hand, the method using the island drive method is unlikely to reject the correct hypothesis at the early stage of processing and can be expected to have a high recognition rate. However, if one word candidate in a spoken sentence is used as the initial hypothesis, the target word appears. Since all the syntactical states that are set are hypotheses, the amount of processing required for the verification increases. Moreover, since the initial hypothesis is determined after detecting all the word candidates in the generated sentence, real-time processing cannot be performed.

The conventional techniques have the above problems.

The problem to be solved by the present invention is to enable a continuous speech recognition apparatus using the syntax processing apparatus of the present invention to reduce the processing amount without lowering the sentence recognition rate and to enable real-time processing as compared with the prior art. .

[Means for Solving the Problems]

In order to achieve the above object, the syntax processing device for continuous speech recognition of the present invention is characterized by including the following means.

In a syntax processing device that receives input word candidates from an acoustic processing device that acoustically processes input speech with a continuous speech recognition device and obtains a speech recognition result by applying syntax rules, acoustic features of input speech and standard pattern acoustics Threshold value determining means for comparing the acoustic distance value of the input word candidate, which is the difference with the dynamic feature amount, with the threshold value, and the output, that is, the word candidate, which is less than or equal to the threshold value by the threshold value determining means, and the syntax rule in the syntax state transition network format, which will be described later. Syntax determining means for determining connection of a word candidate to a word string candidate read out from the word string candidate storing means, and a state transition network format represented by a certain word name, a word name that can be connected before it, and a transition state at the time of connection The reverse lookup syntax rule storage means for storing the input word candidates, the word candidate storage means for storing the input word candidates whose threshold value determination means is equal to or more than the threshold value and is unlikely to be connected to the word string candidates, and the syntax determination means. A backward-looking syntax for determining connection to a word string candidate by applying a rule in the reverse-lookup syntax rule storage means to a word candidate determined to have no connection to a string candidate and a word candidate from the word candidate storage means. A word string candidate storing means for storing word string candidates generated by the syntax judging means or the backward syntax judging means, and a word string candidate having a minimum total acoustic distance value at the end of the acoustic processing of the input voice. Is output.

[Work]

The continuous speech recognition syntax processor configured as described above is
First, it is determined whether the input word candidate obtained by acoustically processing the input voice by the threshold value determination means is included in the voice recognition result, and the input word candidate having a low possibility is temporarily stored.

Next, the syntax determination unit determines whether or not the input word candidate determined to be likely to be included in the speech recognition result by the threshold value determination unit can be connected after the word string candidate by applying the syntax rule, The connectable input word candidates are additionally stored as new word string candidates formed by connecting to the word string candidates.

Further, the backward syntax determining means can connect as a word following the word candidate stored in the input word candidate determined to be unable to connect to the word string candidate by the syntax determining means, and the word candidate can be connected after the word string candidate. Whether or not it is determined by applying the reverse-lookup syntax rule, and the connectable input word candidates are additionally stored as new word string candidates formed by connecting the word candidates and the word string candidates.

Finally, when the acoustic processing end signal is input by the recognition result determination means, the end position of the word string candidate generated by the above-described syntax processing is near the end position of the input voice and the total acoustic distance value is the minimum. Output as.

〔Example〕

FIG. 1 is a functional block diagram of a syntax processing device for continuous speech recognition showing an embodiment of the present invention.

In the figure, (1) is a threshold judgment of the acoustic distance value of the input word candidate (11) obtained by acoustically processing the input speech, and the input word candidate (11) is judged as a syntax or as a word candidate. Threshold decision means for deciding whether to store once, (2) is the end of the word string candidate (22) whose starting end position of the input word candidate (11) from this threshold decision means is read from the word string candidate storage means (5) A syntax rule (33) indicating that the word name w of the input word candidate (11) is accepted in the final state p of the word string candidate (22) adjacent to the position is searched from the syntax rule storage means (6). When possible, a threshold judgment of the total acoustic distance value of the new word string candidate (22) and the input word candidate (11) is performed, and additional memory is stored as the new word string candidate (22). Or there is no word string candidate that can be connected The syntax determining means for determining further whether to backward syntax determined input word candidates (11),
In (3), the start position of the input word candidate (11) from the syntax determining means is adjacent to the end position of the word candidate (55) read from the word candidate storing means (8), and the word candidate (55).
Is adjacent to the end position of the word string candidate (22) read from the word string candidate storage means (5), and the word name w of the word candidate (55) is in the final state s of the word string candidate (22). 'And the subsequent input word candidate (11) is connected to the word name w, the reverse lookup syntax rule (44) can be retrieved from the reverse lookup syntax rule storage means (7), and the word string candidate (22) Whether or not to perform a threshold judgment of the total acoustic distance value of a new word string candidate formed by connecting the word candidate (55) and the input word candidate (11) and to additionally store it as a new word string candidate (22) The backward-syntax determining means (4) determines the end position of the word string candidate generated by the above-described syntax processing when all the input word candidates (11) are input and the acoustic processing end signal (66) is input. Is stored near the end position of the input speech and the total acoustic distance value is the minimum Recognition result judging means for outputting the read speech recognition result (77) from the means (5), (5)
Shows the word string candidates (22) generated by the syntactic process as shown in FIG. 2 for each word address, word name, word acoustic distance value, word string total acoustic distance value, start position, end position, Word string candidate storage means for storing the number of words from the beginning of the sentence, the state, and the configuration of the address in which the preceding word of each word is stored,
(6) is a state p and a word name w that can be accepted from that state
And the state q to which the word transits when the word is accepted are pre-stored in the syntactic rule (33) described in the syntactic state transition network as shown in FIG. 3, and the state pi and the word name wij as shown in FIG. Is used as a key, the syntax rule storage means is configured to search for the transition state qij by comparing with the first two words, (7) is a word name w and a word name w that can be connected before the word name w The states s, p, and q that transit when connected to
As shown in the figure, the reverse lookup syntax rule (44) described in the syntax state transition network is stored in advance, and the word name w'ij can be connected in front of a certain word name wi as shown in FIG. sij,
Reverse lookup syntax rule storage means configured to be able to search pij and qij backward in the direction opposite to the time axis, and (8) is a word candidate (11) that is unlikely to belong to a word string candidate (22).
Is a word candidate storage means for storing.

In the continuous speech recognition syntax processing apparatus of the above-described embodiment, first, an input that includes a word name, a start / end position, and acoustic distance value generated by an acoustic processing apparatus that acoustically processes an input speech by the continuous speech recognition apparatus. Receive word candidates (11).

Next, when the acoustic distance value of the input word candidate (11) is less than or equal to the threshold value by the threshold value judging means (1), it is considered that the input word candidate (11) is likely to be included in the speech recognition result (77), and the syntax judging means. The input word candidate (11) is transmitted to (2). On the other hand, if the threshold value is exceeded, it is considered that the similar possibility is low, and the input word candidate (11) is temporarily stored in the word candidate storage means (8).
Further, the start position of the input word candidate (11) in the syntax judging means (2) is adjacent to the end position of the word string candidate (22) read from the word string candidate storing means (5), and further the word string candidate (22).
When the syntax rule (33) indicating that the word name wij of the input word candidate (11) is accepted in the final state pi of the can be retrieved from the syntax rule storage means (6), the word string candidate (22) and the input word When the total acoustic distance value of a new word string candidate formed by connecting the candidates (11) is less than or equal to the threshold value, it is determined that the input word candidate (11) can be connected to the read word string candidate (22), and the word string candidate It is additionally stored in the storage means (5) as a new word string candidate (22). On the other hand, when there is no word string candidate that can connect the input word candidate (11) beyond the threshold value, the input word candidate (11) is transmitted to the backward syntax determining means (3).

Further, the starting position of the input word candidate (11) in the backward syntax determining means (3) is adjacent to the ending position of the word candidate (55) read from the word candidate storing means (8), and the word candidate (55).
Is adjacent to the end position of the word string candidate (22) read from the word string candidate storage means (5), and the word name w of the word candidate (55) is in the final state sij of the word string candidate (22). ′ I
When a reverse lookup syntax rule (44) indicating that j and the word name wi of the preceding input word candidate (11) are connected can be retrieved from the reverse lookup syntax rule storage means (7), the word string candidate (2
2) The word candidate (55) and the input word candidate (11) are connected, and when the total acoustic distance value of the new word string candidate is less than or equal to the threshold value, the word candidate (55) read by the input word candidate (11). Then, it is determined that the word string candidate (22) can be connected, and additionally stored in the word string candidate storage means (5) as a new word string candidate (22).

When all the input word candidates (11) have been subjected to the above threshold value judgment, syntax judgment and backward syntax judgment processing and all input has been completed, a sound processing end signal (66) is received from the sound processing device.

Finally, the end position of the word string candidate (22) generated by the above processing from the word string candidate storage means (5) in the recognition result judging means (4) is near the end position of the input voice, and the total acoustic distance value is the smallest. The object is read and output as the speech recognition result (77).

〔The invention's effect〕

The present invention is configured as described above, and does not perform the syntax judgment on all the input word candidates as in the prior art, but only the input word candidates that have a small acoustic distance value and are likely to be included in the speech recognition result. If it is not possible to make a decision, remember the ones that are unlikely to be stored once, and if the acoustic distance value is small and it is unlikely that you can connect to a word string candidate, go back to the stored word candidates and make a syntax decision. Therefore, the enormous amount of processing, which was a problem in the conventional method using the forward method, can be reduced without rejecting the correct hypothesis. In addition, even though there is some backward processing, processing is basically performed in the time axis direction, so processing can only be started after the occurrence of the problem, which was a problem with the conventional island driving method. It can also be resolved.

Therefore, the continuous speech recognition apparatus using the syntax processing apparatus of the present invention has an effect of reducing the processing amount without lowering the sentence recognition rate and enabling real-time processing as compared with the conventional apparatus.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a syntax processing device for continuous speech recognition showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of a memory configuration of a word string candidate storage means of FIG. 1, FIG. 3 and FIG. FIG. 4 is a diagram showing a syntax state transition network describing the syntax rules of the syntax rule storage means of FIG. 1 and an example of its storage configuration, and FIGS. 5 and 6 are the reverse of the reverse lookup syntax rule storage means of FIG. FIG. 7 is a diagram showing a syntax state transition network for describing a pull syntax rule and its storage configuration example. FIG. 7 is a functional block diagram of a syntax processing device for continuous speech recognition showing a conventional example, FIG. 8, FIG. 9, FIG. FIG. 8 is a diagram showing a method using a forward method, a beam search method, and an island driving method in the conventional example of FIG. 7. In the figure, (1) is a threshold judgment means, (2) is a syntax judgment means, (3) is a backward syntax judgment means, (4) is a recognition result judgment means, (5) is a word string candidate storage means, and (6) is Syntax rule storage means, (7) is reverse lookup syntax rule storage means, (8)
Is a word candidate storage means, (11) is an input word candidate, (22) is a word string candidate, (33) is a syntax rule, (44) is a reverse syntax rule, (55) is a word candidate, and (66) is a sound. Processing end signal,
(77) is the voice recognition result. In the drawings, the same reference numerals indicate the same or corresponding parts.

Claims (1)

[Claims] 1. A threshold determination means for comparing an acoustic distance value of an input word candidate, which is a difference between an acoustic feature amount of an input voice and an acoustic feature amount of a standard pattern, with a threshold value; A syntax determination means for determining the connection of the word candidate to the word string candidate read out from the word string candidate storage means, which will be described later, and a word that can be connected before a certain word name A reverse lookup syntax rule storage unit that stores a name and a transition state at the time of connection in a state transition network format, and an input word candidate that is less than the threshold value determination unit and is unlikely to be connected to a word string candidate. A word string by applying the rule in the reverse-lookup syntax rule storage means to the word candidate storage means, the word candidate determined by the syntax determination means to have no connection to the word string candidate, and the word from the word candidate storage means. Connection to the candidate And a word sequence candidate storage unit that stores a word sequence candidate generated by the syntax determination unit or the backward syntax determination unit, and when the acoustic processing of the input voice is completed, the total acoustic distance value is A syntax processing device for continuous speech recognition, which outputs the smallest word string candidate.