JPH09212190A - Speech recognition device and sentence recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To greatly reduce the workload of grammar generation corresponding to speech recognition by employing a mechanism which automatically allow a partial sentence during a speech recognizing process. SOLUTION: A recognizing circuit 103 collates LPC cepstrum coefficients Cj(t') of acoustic feature parameters, frame by frame, according to a recognition grammar rule and feature information on a speech and supplies collation results obtained so far to a collation halfway result holding memory circuit 104. This circuit 104 temporarily stores and manages the partial sentence word series PS of a partial sentence (m) and its partial sentence similarity PV, and the sentence word series S of a sentence (n) and its sentence similarity V, and retrieval control over maximum similarity is performed under the instruction control of a partial sentence decision circuit 105. The circuit 105 retrieves the collation result of the circuit 104 when speech termination is established. When the maximum value of the partial sentence similarity is larger than the maximum value of sentence collation similarity by a certain extent, the partial sentence giving the maximum similarity is outputted as a speech recognition result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and a sentence recognition device, and can be applied to recognition when acoustic signals or character string information of at least one sentence are continuously input.

[0002]

2. Description of the Related Art In a speech recognition apparatus that allows ordinary sentence speech, it is necessary to describe a sentence to be recognized in advance based on a predetermined grammar. This grammar description rule generally uses a form called a finite state automaton or a context-free grammar.

[0003]

However, the grammatical forms called finite state automata and context-free grammars are
It is a format that expresses the rules for connecting words from the beginning to the end of a sentence. Therefore, of course, in a sentence that matches this grammatical rule, the words from the beginning of the sentence to the end of the sentence must be correctly connected.

Generally, in a sentence written in a text, such a problem does not cause any particular problem, but in continuous speech input, it is rather less that a complete sentence is uttered. Abbreviations, noun stops, and multiple insertions of interjections (touch words) are occurring. This means that when using a grammar intended for general texts for speech recognition, it is necessary to allow some sub-sentences that do not start from the beginning of a sentence and reach the end of the sentence.

Then, a grammar rule allowing this sub-sentence should be created in advance and applied to speech recognition. However, describing a grammar in which all such sub-sentences are regarded as sentences is By nature, it was a labor-intensive task that was prone to human error.

From the above, it is possible to improve the voice recognition rate by introducing a mechanism for automatically allowing a partial sentence during the voice recognition process to remarkably reduce the amount of work for creating a grammar corresponding to the voice recognition. There is a demand for a voice recognition device, and a sentence recognition device for recognizing a sentence or a partial sentence when at least one sentence of character string information is input.

[0007]

Therefore, the invention of claim 1 is a speech recognition apparatus which receives at least one sentence of speech continuously and outputs a speech recognition result. It solves the problem.

That is, the speech recognition apparatus according to claim 1 is
(1) Acoustic feature information is extracted from the input acoustic signal for each predetermined frame, and the acoustic feature information is collated for each predetermined frame based on the voice feature information stored in advance and a predetermined grammar rule, The "input acoustic analysis / matching means" for temporarily storing the matching result, and (2) the matching result for the partial sentence and the matching result for at least one sentence are obtained from the matching result for each frame, and the matching for the partial sentence is performed. The result is compared with the matching result for the one sentence, and a partial sentence or one sentence of one matching result having a high similarity from the above two matching results is output as a speech recognition result. Means ".

By adopting such a configuration, the acoustic feature information is extracted from the input acoustic signal, and is collated with the standard voice feature information stored in advance, and at predetermined frames according to a predetermined grammatical rule. Is compared with the acoustic feature information described above, the comparison result is stored, the comparison result as a partial sentence is compared with the comparison result as one sentence, and the one having a higher similarity degree from the viewpoint of similarity degree is compared. It is possible to output the speech recognition result for the input sound from either the part or one sentence, and even if the similarity as a sentence is low, the partial sentence is recognized with a high recognition rate if the similarity as a sentence is high. It can be output.

According to a fourth aspect of the present invention, in a sentence recognition device which receives character string information of at least one sentence and outputs a sentence recognition result, the above problem is solved by the following characteristic configuration. .

That is, the sentence recognition device according to the invention of claim 4 is: (1) The input character string information is predetermined based on the word dictionary information stored in advance from the input character string information and a predetermined grammar rule. "Input character string information collating means" that collates each long character string and temporarily stores this collation result,
(2) A collation result for the partial sentence and a collation result for at least one sentence are obtained from the collation result for each character string having the predetermined length, and the collation result for the partial sentence and the collation result for the one sentence are compared, It is provided with "partial sentence / sentence similarity determination means" for outputting a partial sentence or one sentence of one matching result having a high similarity from the two matching results as a sentence recognition result.

By adopting such a configuration, the input character string information and the word dictionary information are used to collate the input character string as a word, and according to a predetermined grammatical rule, collation as a partial sentence and a sentence is performed. From these collation results, it is possible to output the sentence recognition result for the input character string information for either the part with higher similarity or one sentence,
Even if the similarity as one sentence is low, if the similarity as a partial sentence is high, the partial sentence can be recognized and output with a high recognition rate.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, a preferred embodiment of the present invention will be described with reference to the drawings. [First Embodiment]: First, in the speech recognition apparatus according to the first embodiment of the present invention, (1) an interim result of matching between a speech signal and a recognition grammar at the time of inputting speech of a speech recognition target. (2) The maximum value of the sentence matching similarity at the time when the voice input is completed is compared with the maximum value of the matching similarity of the partial sentence, and (3) the maximum value of the partial sentence similarity is the sentence. When the maximum value of the matching similarity is greater than a certain value, the partial sentence with the maximum similarity is set as the recognition result, and the maximum value of the partial sentence similarity is not more than a certain value than the maximum value of the sentence matching similarity. In the sentence matching similarity is provided with a determination circuit that uses a sentence as a recognition result.

FIG. 1 is a functional block diagram of the voice recognition device according to the first embodiment. In FIG. 1, the voice recognition device includes a sound analysis circuit 101 and a voice section detection circuit 102.
A recognition circuit 103 and a collation intermediate result holding memory circuit 1
04 and a sub-sentence determination circuit 105.

The acoustic analysis circuit 101 performs A (analog) / D (digital) conversion of an input acoustic signal from an analog signal to a digital signal, performs acoustic characteristic analysis processing at each short time called a frame, and outputs the acoustic characteristic parameter 101S. Is calculated. This frame is, for example, about 10 ms to 30 ms. In addition, the above A / D conversion is
For example, the sampling frequency is, for example, 8 to 16 KHz.
Degree.

The acoustic analysis circuit 101 is, specifically,
As the acoustic feature parameter 101S, for example, LPC
(Linear Prediction Coding: Liner Predictio
nCording) cepstrum coefficient Cj (t ′) is used. Here, j = 0, 1, ..., J, where J is the cepstrum order, t ′ is a frame number, and the start time of the voice recognition device is t ′ = 0.

This LPC cepstrum coefficient Cj (t ')
Is given to the voice section detection circuit 102 and the recognition circuit 103 as the acoustic feature parameter 101S.

The voice section detection circuit 102 sets the time when the voice starts to be uttered from the acoustic feature parameter sequence for each frame, that is, this time as the voice start signal ts, and the time when the utterance ends, that is, this time. Voice termination signal te
And the recognition circuit 103 and the partial sentence determination circuit 10
5 and give.

Specifically, the voice section detection circuit 102 detects the voice start signal ts from the change in the LPC cepstrum coefficient Cj (t '). Regarding the detection of the voice start signal ts, in order to make the explanation easy to understand,
The relative frame number from the voice start signal ts is represented by t,
When t '= ts, t = 0.

That is, in order to make the explanation easy to understand, the relative frame number t = t in the voice section detection circuit 102.
′ −ts, where T is the voice section, T = te−t
s + 1, and the voice termination signal te becomes te = T-1.

The recognition circuit 103 receives the acoustic feature parameter 1 given from the acoustic analysis circuit 101 based on the previously stored recognition grammar rules and standard feature information of voice.
The LPC cepstrum coefficient Cj (t ′) of 01S is collated for each frame, and the collation result up to that point is given to the collation intermediate result holding memory circuit 104.

As the recognition grammar rule, for example,
It is preferable to provide various grammar rules such as finite state automaton, context-free grammar, phrase structure rule, general grammar, and regular grammar.

This recognition circuit 103 is, specifically,
The C cepstrum coefficient Cj (t ') is changed from t = 0 to t = T-
The partial sentence word series PS (m, t) of the partial sentence m up to the frame number t and the partial sentence similarity PV
(M, t), the sentence word sequence S (n, t) of the sentence n, and the sentence similarity V (n, t) are obtained and given to the collation intermediate result holding memory circuit 104. Further, when the matching of one sentence is completed, the matching completion information 103S is generated and output and given to the partial sentence determination circuit 105.

The collation intermediate result holding memory circuit 104 stores the partial sentence word sequence PS of the partial sentence m from the recognition circuit 103.
(M, t) and its partial sentence similarity PV (m, t), and the sentence n
Sentence word sequence S (n, t) and its sentence similarity V (n, t)
Are temporarily stored and managed, and the maximum similarity is searched and controlled by the instruction control from the sub-sentence determination circuit 105.

The sub-sentence determining circuit 105 determines when the voice end is established by the determination from the information of the voice start signal ts and the information of the voice end signal te given from the voice section detecting circuit 102, or from the recognition circuit 103. According to the verification completion information 103S of
The matching result of 04 is searched. That is, the maximum similarity between the partial sentence similarity and the sentence similarity is searched.

These maximum similarities are, for example, the maximum partial sentence similarity MAX PV (m ', T-1) and the maximum sentence similarity MAX V (n', T-1).

If the maximum value of the partial sentence similarity is larger than the maximum value of the sentence matching similarity by a certain amount or more, the partial sentence giving the maximum similarity is output as the speech recognition result. Further, when the maximum value of the partial sentence similarity is not larger than the maximum value of the sentence matching similarity by a certain amount or more, the sentence which can obtain the maximum value among the sentence matching similarities is output as the voice recognition result. .

The sub-sentence judgment circuit 105 is, for example, specifically, the collation completion information 103S from the recognition circuit 103.
Is given, the maximum value MAX PV (m of the similarity PV (m, T-1) of the partial sentence m of the voice termination frame T-1 is given.
', T-1) and the maximum value MAX V (n', T-1) of the similarity V (n, T-1) of the sentence n are searched. Here, m ′ is the number of the partial sentence that gives the maximum value of the similarity, and n ′ is the number of the sentence that gives the maximum value of the similarity.

The sub-sentence determination circuit 105 searches for the maximum value of the degree of similarity to find the maximum sub-sentence MAX PV.
The value of (m ', T-1) -SH is the maximum sentence similarity MAX.
When it is larger than V (n ′, T−1), the partial sentence m ′ that gives the maximum similarity is output as the speech recognition result. In addition, the maximum similarity of partial sentences MAX PV (m ', T-1)-
The value of SH is the maximum sentence similarity MAX V (n ', T-1).
If it is not greater than, the maximum value n ′ in the sentence matching similarity
The sentence that gives is output as the speech recognition result.

(Operation of the voice recognition device): Next, the operation of the voice recognition device of FIG. 1 will be described. First, the recognition circuit 10 in advance
It is assumed that standard voice characteristic information is stored in 3 and a predetermined recognition grammar rule is also stored therein.
Next, when the acoustic signal is given to the acoustic analysis circuit 101,
The acoustic feature parameter 101S such as the LPC cepstrum coefficient Cj (t ') is obtained by this circuit and is given to the voice section detection circuit 102 and the recognition circuit 103. This acoustic feature parameter 101S is used for the voice section detection circuit 102.
To the recognition circuit 103 and the partial sentence determination circuit 105, the circuit 102 detects the voice start signal ts of the utterance and the voice end signal te which is the end of the utterance. .

On the other hand, the acoustic feature parameter 101S is also given to the recognition circuit 103. In this circuit 103, based on the recognition grammar rules stored in advance and the standard feature information of the voice, the acoustic analysis circuit 101 LPC cepstrum coefficient Cj of given acoustic feature parameter 101S
(T ′) is collated for each frame, and the collation result up to that point is given to the collation midway result holding memory circuit 104. That is, the LPC cepstrum coefficient Cj (t
′) Are collated in order from t = 0 to t = T−1, and the partial sentence word series PS (m, of the partial sentence m up to the frame number t
t), the partial sentence similarity PV (m, t), the sentence word series S (n, t) of the sentence n, and the sentence similarity V (n, t) are obtained, and the collation intermediate result is obtained. It is temporarily stored in the holding memory circuit 104.

Partial sentence word sequence PS of these partial sentences m
(M, t), the partial sentence similarity PV (m, t), the sentence word sequence S (n, t) of the sentence n, and the sentence similarity V (n, t).
The collation result such as t) is stored in the interim collation result holding memory circuit 1
When it is given to 04, it is temporarily stored and managed, and the maximum similarity is searched and controlled by the instruction control from the sub-sentence determination circuit 105.

Next, the collation completion information 1 from the recognition circuit 103.
03S or the information of the voice termination signal te is the partial sentence determination circuit 1
When it is given to 05, this circuit 105 searches the collation result of the collation intermediate result holding memory circuit 104 at the time when the voice termination is established. That is, the maximum similarity between the partial sentence similarity and the sentence similarity is searched.

That is, specifically, for example, the recognition circuit 1
When the collation completion information 103S is given from 03, the similarity PV (m, T- of the partial sentence m of the voice termination frame T-1 is given.
1) maximum value MAX PV (m ', T-1) and maximum value MAX V (n', T) of similarity V (n, T-1) of sentence n
-1) and are searched.

Then, by the maximum value search of the above-mentioned similarity degree, the maximum similarity degree of partial sentences MAX PV (m ', T-1)-
The value of SH is the maximum sentence similarity MAX V (n ', T-1).
If it is determined that the maximum similarity is greater than the above, the partial sentence m ′ that gives the maximum similarity is output as the voice recognition result. If it is determined that the value of the maximum partial sentence similarity MAX PV (m ′, T−1) −SH is not greater than the maximum partial sentence similarity MAX V (n ′, T−1), sentence matching is performed. The sentence giving the maximum value n'in the similarities is output as the voice recognition result.

(Effects of the first embodiment of the present invention):
According to the above-described first embodiment of the present invention, the similarity between partial sentences and the similarity between sentences are compared and determined after the voice recognition processing,
When the degree of similarity between sub-sentences is higher than the degree of sentence similarity, by using the result of the sub-sentence as the recognition result, it is possible to eliminate the sub-sentence correspondence work that is a problem when creating a grammar for speech recognition. The work efficiency can be significantly increased.

More specifically, a speech recognition apparatus that incorporates a mechanism for automatically allowing partial sentences during the speech recognition process to significantly reduce the amount of work required to create a grammar corresponding to speech recognition and improve the speech recognition rate of partial sentences. Can be realized.

[Second Embodiment]: Next, the second embodiment of the present invention
In the voice recognition device of the embodiment, (1) the voice input data and the recognition grammar are collated from the beginning of the sentence, and the maximum value of the sentence matching similarity and the maximum value of the partial sentence matching similarity from the beginning of the sentence are obtained. (2) Further, the voice input data and the recognition grammar are collated from the end of the sentence to obtain the maximum value of the partial sentence collation similarity from the end of the sentence. (3) When the maximum value of the partial sentence similarity from the beginning of each sentence and the end of the sentence is larger than the maximum value of the sentence matching similarity by a certain amount or more, the partial sentence giving the maximum similarity is set as the recognition result. , And a determination circuit that uses a sentence having the maximum value among the sentence matching similarities as a recognition result.

FIG. 2 is a functional block diagram of a voice recognition device according to the second embodiment of the present invention. In FIG. 2, the voice recognition device includes an acoustic analysis circuit 101, a voice section detection circuit 102, a voice dictionary circuit 201, and a forward direction recognition circuit 20.
2, a backward direction recognition circuit 203, a collation intermediate result holding memory circuit 104A, and a partial sentence determination circuit 105A.

The speech recognition apparatus of the second embodiment is characterized by a voice dictionary circuit 201, a forward direction recognition circuit 202, a backward direction recognition circuit 203, a collation intermediate result holding memory circuit 104A, Sub-sentence judgment circuit 105A
And The other acoustic analysis circuit 101 and voice section detection circuit 102 have the same functions as those in the first embodiment.

The voice dictionary circuit 201 stores standard voice feature information in advance. This voice feature information is used as the forward direction recognition circuit 202 and the backward direction recognition circuit 203.
To give to.

The forward direction recognition circuit 202 is the acoustic analysis circuit 1
When the acoustic feature parameter 101S is given from 01,
Based on the recognition grammar rules stored in advance, the voice feature information stored in the voice dictionary circuit 201 is used to collate the acoustic feature parameters 101S in the input order in the forward direction. It is stored in the collation midway result holding memory circuit 104A.

That is, the LPC cepstrum coefficient Cj (t), which is the acoustic feature parameter 101S, is changed from t = 0 to t = T.
Forward matching is performed from the beginning of the sentence up to -1, and the word sequence FS (m) of the forward partial sentence m up to the frame number t and the similarity F thereof.
V (m), the word sequence S (n) of the sentence n, and its similarity V
(N) is obtained, and the interim collation result holding memory circuit 104A is obtained.
To give to.

The backward direction recognition circuit 203 is the acoustic analysis circuit 1
When the acoustic feature parameter 101S is given from 01,
On the basis of the recognition grammar rules stored in advance, the voice feature information stored in the voice dictionary circuit 201 is used to reverse the order in which the acoustic feature parameters 101S are input,
The backward collation result from the voice end te to the voice start end ts is collated, and the backward collation result is held in the collation intermediate result holding memory circuit 10.
It is stored in 4A.

That is, the LPC cepstrum coefficient Cj (t), which is the acoustic feature parameter 101S, is calculated from t = T-1 to t.
From the end of the sentence to the beginning of the sentence until = 0, the word sequence BS (S) of the backward partial sentence S up to frame 0 and its similarity BV
(S) is given to the mid-verification result holding memory circuit 104A.

The mid-verification result holding memory circuit 104A
The word sequence FS (m) of the forward partial sentence m from the forward recognition circuit 202 and its similarity FV (m), the word sequence S (n) of the sentence n and its similarity V (n), and the reverse direction. Recognition circuit 20
The partial sentence determination circuit 10 stores and manages the word sequence BS (S) of the backward partial sentence S from 3 and its similarity BV (S).
From 5A, the maximum similarity of the partial sentence and the maximum similarity of the sentence are search-controlled.

The sub-sentence determination circuit 105A searches the collation intermediate result holding memory circuit 104A and selects the maximum value of the sub-sentence similarity and the maximum value of the sentence collation similarity. When the maximum value of the partial sentence similarity is larger than the maximum value of the sentence matching similarity by a certain amount or more, the partial sentence having the maximum similarity is output as the voice recognition result. Further, when the maximum value of the partial sentence similarity is not larger than the maximum value of the sentence matching similarity by a certain amount or more, the sentence which obtains the maximum value among the sentence matching similarities is output as the voice recognition result.

That is, the sub-sentence determining circuit 105A determines the similarity FV (m) of the forward sub-sentence m, the maximum value MAX PV in the similarity BV (S) of the backward sub-sentence S, and the sentence n. The maximum value MAX V of the similarity V (n) is searched.

As a result, if the maximum similarity M of the partial sentence is
When the value of AX PV-SH (which is a threshold value greater than zero indicating the priority of a sentence) is larger than the maximum similarity MAX V of a sentence, a forward portion that gives the maximum similarity MAX PV of a partial sentence The sentence m or the backward partial sentence S is output as the voice recognition result. In addition, the maximum similarity M of partial sentences
When the value of AX PV-SH is not larger than the maximum similarity MAX V of the sentence, the sentence that can obtain the maximum similarity is output as the speech recognition result.

(Operation of Speech Recognition Device) Next, referring to FIG.
The operation of the voice recognition device will be described. First, it is assumed that a predetermined recognition grammar rule is stored in advance in the forward direction recognition circuit 202 and the backward direction recognition circuit 203. Furthermore, it is assumed that the voice dictionary circuit 201 stores standard feature information of voice. Next, the acoustic signal is the acoustic analysis circuit 1
01, the acoustic characteristic parameter 101S such as the LPC cepstrum coefficient Cj (t ′) is obtained by this circuit, and the voice section detecting circuit 102 and the forward direction recognizing circuit 2 are obtained.
02 and the reverse direction recognition circuit 203.

When this acoustic feature parameter 101S is supplied to the voice section detection circuit 102, this circuit 102S.
At the same time, the voice start signal ts of utterance is detected, and the voice end signal te which is the end of utterance is detected and applied to the forward direction recognition circuit 202 and the backward direction recognition circuit 203.

The acoustic feature parameter 101S, the voice start signal ts of the utterance, and the voice end signal te are the forward direction recognition circuit 2
When given to No. 02, the acoustic feature parameters 101S are collated in the forward order in the input order using the speech feature information stored in the voice dictionary circuit 201 based on the recognition grammar rules stored in advance. The forward verification result is given to the mid-verification result holding memory circuit 104A.

Specifically, the acoustic feature parameter 101S
The LPC cepstrum coefficient Cj (t) is collated from t = 0 to t = T−1 in the forward direction from the beginning of the sentence, and the word sequence FS (m) of the forward direction partial sentence m up to the frame number t and its similarity. FV (m), the word sequence S (n) of the sentence n, and the similarity V (n) thereof are obtained and given to the collation intermediate result holding memory circuit 104A.

Further, when the acoustic feature parameter 101S, the voice start signal ts of the utterance, and the voice end signal te are given to the backward direction recognition circuit 203, the voice dictionary circuit 201 receives the voice grammar circuit 201 based on the recognition grammar rules stored in advance. The acoustic feature parameter 101 is also stored using the stored voice feature information.
Contrary to the order in which S is input, collation is performed in the reverse direction from the voice end signal te to the voice start signal ts, and the backward collation result is given to the collation midway result holding memory circuit 104A.

Specifically, the acoustic feature parameter 101S
LPC cepstrum coefficient Cj (t) is t = T-1
From the end of the sentence to the beginning of the sentence from 0 to t = 0, and the frame 0
Up to and including the word sequence BS (S) of the backward partial sentence S and its similarity BV (S) are in the middle of collation result holding memory circuit 104A.
Given to.

In the collation intermediate result holding memory circuit 104A, the word sequence FS (m) of the forward partial sentence m from the forward recognition circuit 202 and its similarity FV (m) and the word sequence S (n) of the sentence n. And its similarity V (n), the word sequence BS (S) of the backward partial sentence S from the backward recognition circuit 203, and its similarity BV (S) are stored and managed, and the partial sentence determination circuit 105A outputs them. The maximum similarity of the partial sentence and the maximum similarity of the sentence are search-controlled.

By the search by the partial sentence determination circuit 105A, the maximum value of the partial sentence similarity and the maximum value of the sentence matching similarity are selected. Then, when the maximum value of the partial sentence similarity is larger than the maximum value of the sentence matching similarity by a certain amount or more,
The partial sentence that has the maximum similarity is output as the voice recognition result. When the maximum value of the partial sentence similarity is not higher than the maximum value of the sentence matching similarity by a certain amount or more, the sentence that obtains the maximum value among the sentence matching similarities is output as the voice recognition result.

That is, in this sub-sentence determination circuit 105A,
The similarity FV (m) of the forward sub-sentence m and the backward sub-sentence S
The maximum value MAX PV of the similarity BV (S) and the maximum value MAX V of the similarity V (n) of the sentence n are searched.

As a result, the maximum similarity MAX of partial sentences is obtained.
When the value of PV-SH (a threshold value greater than zero indicating the priority of a sentence) is larger than the maximum similarity MAX V of a sentence, a forward partial sentence that gives the maximum similarity MAX PV of a partial sentence The m or backward partial sentence S is output as the voice recognition result. Also, the maximum similarity MA of partial sentences
When the value of X PV-SH is not larger than the maximum similarity MAX V of the sentence, the sentence that can obtain the maximum similarity is output as the speech recognition result.

(Effects of the second embodiment of the present invention):
According to the speech recognition apparatus of the second embodiment of the present invention described above, the similarity between the forward partial sentences and the similarity between the backward partial sentences and the sentences are compared and determined to determine the similarity of the partial sentences. When the degree is higher than the sentence similarity, by using the result of the sub-sentence as the recognition result, it is possible to eliminate the sub-sentence correspondence work which is a problem when creating the grammar for speech recognition, and the grammar creation work efficiency is significantly improved. It can be increased.

Further, in the backward partial sentence, it is possible to easily remove a prefix frequently occurring during utterance, for example, words such as "er" and "an", from the recognition target, and to obtain a substantial recognition performance. It can also contribute to improvement.

Therefore, a mechanism for automatically permitting partial sentences during the speech recognition process is introduced, and the amount of work for creating a grammar corresponding to speech recognition is significantly reduced, and a speech recognition apparatus for improving the speech recognition rate of partial sentences is realized. be able to.

(Other Embodiments): (1) In the above embodiment, a sentence is generally a minimum unit, and a phrase formed of several words, or a phrase is a sentence. When a shape is formed, it is formed in a hierarchy such as a clause, so a sub-sentence is one clause, one phrase,
It may be any of the minimum unit words. In addition, one sentence can be applied even when clauses are connected to each other.

(2) In addition to the LPC cepstrum,
Features can be captured by the LPC mel cepstrum or the simplest power spectrum (spectral formant), and the dynamic parameter delta cepstrum (the amount of time change of the cepstrum is expressed by time difference). It is also preferable to capture the feature with).

(3) Further, in the above speech recognition, continuous D
P (Dynamic Programming: DynamicProgramming
g: Non-linear expansion / contraction pattern) Partial sentence by matching,
It is also possible to collate the sentences. In addition, HMM (Hidden Mark) which is statistical / probabilistic pattern recognition is used.
ov Model: Hidden Markov Model) method is also preferably used.

(4) Furthermore, the present invention can be applied not only to recognition of partial sentences and sentences by the voice recognition device, but also to a sentence recognition device which recognizes a sentence when one or more character strings are input. Specifically, for example, a character is read by an OCR (character reading device) or the like, and the read character string is processed by a recognition circuit or a partial sentence determination circuit using a word dictionary to generate a partial sentence or a single sentence. Can recognize.

(5) It can also be applied to a case where a specific part is searched out from a database in which a sentence is input, and it can also be applied to the field of information retrieval.

[0068]

As described above, according to the first aspect of the invention, in a voice recognition device in which at least one sentence of voice is continuously input and a voice recognition result is output, an audio signal is output from an input audio signal at predetermined frames. An input acoustic analysis that extracts characteristic information, collates the acoustic characteristic information for each predetermined frame based on prestored characteristic information of voice and a predetermined grammar rule, and temporarily stores the collation result. -The matching means and the matching result for the partial sentence and the matching result for at least one sentence are obtained from the matching result for each frame,
A partial sentence / sentence that compares the matching result for the sub-sentence with the matching result for the one sentence and outputs one of the matching partial sentences or one sentence having a high similarity from the two matching results as the speech recognition result. And a similarity determining means.

By adopting such a configuration, the present invention provides
It is possible to implement a speech recognition apparatus that incorporates a mechanism for automatically allowing partial sentences during the speech recognition process, significantly reduces the amount of work for creating a grammar corresponding to speech recognition, and improves the speech recognition rate of partial sentences.

According to a fourth aspect of the present invention, in a sentence recognition device in which character string information of at least one sentence is input and a sentence recognition result is output, predetermined word dictionary information and predetermined word dictionary information are input from the input character string information. Based on the grammar rules of
The input character string information is collated for each character string of a predetermined length, input character string information collating means for temporarily storing the collation result, and the collation result for the partial sentence from the collation result of each character string of the predetermined length, The collation result for at least one sentence is obtained, the collation result for the partial sentence is compared with the collation result for one sentence, and the sub-sentence or one sentence of the one collation result having a high similarity from the two collation results is the sentence recognition result. The sub-sentence / sentence similarity determining means for outputting

By adopting such a configuration, the present invention provides
It is possible to realize a sentence recognition device that recognizes a sentence or a partial sentence when character string information of at least one sentence is input.

[Brief description of drawings]

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

FIG. 2 is a functional configuration diagram of a voice recognition device according to a second embodiment of the present invention.

[Explanation of symbols]

101 ... Acoustic analysis circuit, 101S ... Acoustic feature parameter, 102 ... Voice section detection circuit, 103 ... Recognition circuit, 1
04, 104A ... Collation intermediate result holding memory circuit, 10
5, 105A ... Sub-sentence determination circuit, 201 ... Voice dictionary circuit, 202 ... Forward recognition circuit, 203 ... Reverse recognition circuit, MAX PV (m ', T-1) ... Sub-sentence maximum value, MAX V ( n ′, T−1) ... maximum value of sentence similarity, t
s ... voice start signal, te ... voice end signal, m ... partial sentence,
n ... one sentence, PS (m, t) ... partial sentence word series, PV
(M, t) ... partial sentence similarity, S (n, t) ... sentence word sequence, V (n, t) ... sentence similarity.

Claims (5)

[Claims] 1. A voice recognition device in which voices of at least one sentence are continuously input and a voice recognition result is output, and acoustic feature information is extracted from an input acoustic signal for each predetermined frame, and the voice feature information stored in advance is stored. Based on the characteristic information and a predetermined grammar rule, the acoustic characteristic information is collated for each of the predetermined frames, and an input acoustic analysis / collation means for temporarily storing the collation result, and a portion from the collation result for each frame. The collation result for the sentence and the collation result for at least one sentence are obtained, the collation result for the partial sentence is compared with the collation result for the one sentence, and one collation result having a high similarity from the two collation results is obtained. And a partial sentence / sentence similarity determining unit that outputs the partial sentence or one sentence as a speech recognition result. 2. The speech recognition apparatus according to claim 1, wherein the matching result of the partial sentence and the one sentence includes a similarity degree obtained by matching the word series. 3. The input acoustic analysis / collation means performs the collation in the forward direction in the input start order of the acoustic signal, and stores the forward collation result temporarily, and the acoustic signal. And a backward matching unit that performs the matching in the reverse direction from the input end time to the input start time, and temporarily stores the backward matching result, and the partial sentence / sentence similarity determination means is 3. The voice recognition device according to claim 1, wherein a comparison result for the partial sentence and a comparison result for the one sentence are compared based on a forward matching result and a backward matching result. 4. A sentence recognition device for inputting at least one sentence of character string information and outputting a sentence recognition result, based on word dictionary information stored in advance from input character string information and a predetermined grammatical rule, Input character string information collating means for collating the input character string information for each character string of a predetermined length and temporarily storing the collation result, and a collation result for a partial sentence from the collation result of each character string of the predetermined length. , A matching result for at least one sentence is compared, and the matching result for the partial sentence is compared with the matching result for the one sentence, and a partial sentence or one sentence of one matching result having a high similarity from the above two matching results. A sentence recognition device, comprising: a partial sentence / sentence similarity determination means for outputting as a sentence recognition result. 5. The input character string information collating means performs the collation in the forward direction in the input start order of the input character string information, and a forward collating unit for temporarily storing the forward collation result, The above-mentioned collation is performed in the reverse direction from the input end time of the input character string information to the input start time, and a reverse direction collation unit for temporarily storing the reverse direction collation result is provided, and the sub-sentence / sentence similarity determination is performed. 5. The sentence recognition apparatus according to claim 4, wherein the means compares the matching result for the partial sentence with the matching result for the one sentence from the forward matching result and the backward matching result.