JP2002258891A - Method, device and program for recognizing voice, and recording medium for the program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognizing method for finding semantic contents of speech from inputted voice with high accuracy by preventing probability dispersion due to an unnecessary calculation. SOLUTION: In the method for recognizing voice, an input voice parameter string and semantic representation with respect to the proceeding speech are received; occurrence probability, extracted from a table where the occurrence probability of each semantic representation of the next speech, is recorded is inputted; the likelihood calculation of semantic representation of a semantic representation list stored with semantic representation representing the semantic contents of a speech sentence assumed to be made by a user corresponding to the input voice parameter string is carried out by using each piece of information of an acoustic model, where syllable standard patterns corresponding to voice parameter strings is recorded; and a sentence model, representing conditioned occurrence probability between semantic representation corresponding to the syllable chain standard pattern of the speech sentence assumed to be made by the user, and semantic representation having the maximum likelihood, is outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method and apparatus, a program and a recording medium for inputting human speech, receiving a linguistic meaning of the inputted speech, and outputting a semantic expression, and relates to a product by speech. It is used for processing of orders and the like.

[0002]

2. Description of the Related Art The configuration of a conventional speech recognition apparatus is shown in FIG. For example, FFT (Fast Fourier Trans
A characteristic parameter sequence of an input speech (hereinafter, referred to as an “input speech parameter sequence”) is extracted by a speech analysis unit based on form) analysis and LPC (Linear Predictive Coding) analysis. The input speech parameter sequence is compared with a recognition acoustic model in which the correspondence between the speech parameter sequence and the syllable (acoustic chain pattern) occurrence probability and a recognition language model in which linguistic constraints are defined are recognized and recognized by a speech recognition unit. Generate and output the resulting recognized word string.

A semantic analysis unit inputs a recognized word string output from the speech recognition unit, analyzes the recognized word string using a semantic analysis grammar described in rules, and expresses the result in a semantic expression (for example, a category). Output).

[0004]

In a conventional speech recognition apparatus, a set of recognition word strings generated by a speech recognition unit using a recognition language model and a semantic analysis grammar used by a semantic analysis unit can be analyzed. The sets of word strings were different. As a result, in some cases, the semantic analysis unit cannot analyze the input recognized word string using the semantic analysis grammar. In this case, the recognized word string input to the semantic analysis unit was rejected, and no valid output was made from the semantic analysis unit. In the speech recognition unit, outputting a recognized word string that is rejected in later processing indicates that excessive processing is being performed, and this excessive processing is generated by the efficiency of speech recognition and excessive processing. The data disperses the probabilities and lowers the speech recognition rate.

Further, a method has been proposed in which a grammar for semantic analysis described as a rule (for example, a context-free grammar) is uniformly used as a language model for speech recognition. To narrow down, an enormous number of rules must be described, which is costly and has a problem in its feasibility.

[0006]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention does not dynamically generate a sentence during speech recognition, but uses a unified language for a set of sentences associated with a semantic expression. A method for statistically performing processing corresponding to speech recognition and semantic analysis by using the model as a model and probabilistically predicting the content of the next utterance is provided.

[0007]

FIG. 1 shows the configuration of an embodiment of a speech recognition apparatus according to the present invention. This speech recognition device extracts an input speech parameter sequence from an inputted speech and outputs the linguistic content of the utterance as a semantic expression. The speech recognition device includes a speech analysis unit and a semantic expression analysis unit. The semantic expression analysis unit includes a likelihood calculation unit, an acoustic model storage unit, a sentence model storage unit, a semantic expression list storage unit, and an utterance prediction unit having a next utterance prediction table.

The voice analysis unit extracts an input voice parameter sequence (feature parameter sequence) from the input voice by FFT analysis, LPC analysis, or the like, and outputs it. The likelihood calculation unit receives the input speech parameter sequence and the semantic expression corresponding to the immediately preceding utterance and receives a list of expected occurrence probabilities for all semantic expressions in the received next utterance (referred to as “collection of next utterance meaning expression occurrence probabilities”). Then, using the information stored in the acoustic model and the information stored in the sentence model, a likelihood calculation for obtaining and outputting a semantic expression corresponding to the input speech parameter sequence is executed.

The acoustic model is a syllable standard pattern that is statistically represented by a hidden Markov model or the like created using the same speech parameter sequence (feature parameter) as that used for the input speech parameter sequence. .
That is, syllables (C (consonant) V (vowel) units) are recorded in correspondence with acoustic features in the frequency domain.
(See Seiichi Nakagawa, "Speech Recognition by Probabilistic Model," IEICE (1988).) The semantic expression list is a record of all semantic expressions for utterance sentences assumed to be uttered by the user.

FIG. 2 shows a specific example of the semantic expression list. In this example, the semantic expression is represented by a combination of three values. Instruction types (list display, product order, order end) indicating the utterance of the user, and their arguments (category name “vegetable”, product name “cabbage”, quantity “1”), respectively
It is the information which becomes. (However, "meaning" explains the contents of each semantic expression and is not included in the contents of the list.) The sentence model is a syllable chain pattern of the user's uttered sentence and an utterance about a certain semantic expression. Is recorded, the conditional occurrence probability (a certain utterance sentence is generated when a certain meaning expression is generated) at which the syllable chain pattern is generated when the user performs is performed. That is, the correspondence between the syllable chain pattern and the conditional occurrence probability for each semantic expression is recorded.

FIG. 3 shows a specific example of a sentence model. The syllable chain pattern in this sentence model is described using the same syllable type as the syllable type whose standard pattern is stored in the acoustic model. Here, it is represented using 26 syllable types. The utterance prediction unit inputs the previous utterance semantic expression using a next utterance prediction table that records the correspondence between the previous semantic expression and the appearance probability of each next semantic expression when requested by the likelihood calculation unit. Then, the utterance prediction for outputting the probability of appearance of the meaning expression of the utterance assumed to be the next utterance is executed, and the next utterance meaning expression appearance probability collection is output.

FIG. 4 shows a configuration example of the utterance prediction unit. In the next utterance prediction table, each row has a semantic expression (M1,
.., Mx,..., Mn), and the corresponding semantic expression is stored in the leftmost column of the table. Each column other than the leftmost column has each semantic expression (M1,
.., Mx, ..., Mn), and each number on that column has a condition that, when the semantic expression in the leftmost column appears, each semantic expression appears in the next utterance Probability (P (M1 | M1), ...
・, P (Mn | M1)), ..., (P (M1 | Mn), ..., P (Mn | M
n)).

The utterance prediction processing unit performs the following processing upon receiving the semantic expression of the previous utterance as input. The input semantic expression is checked against each semantic expression in the leftmost column of the next utterance prediction table, and the numerical values (P (M1 | Mx), .., P (Mn | Mx)) are output as a collection of next utterance meaning expression appearance probabilities. FIG. 5 shows a specific example of the next utterance prediction table. In this table, the symbol “*” in the semantic expression indicates that the value in the semantic expression can have any value. This is a measure for efficiently creating a table.

The processing procedure of likelihood calculation by the likelihood calculation unit is shown in FIG.
A description will be given with reference to the flow shown in FIG. 7 and an example of the calculation result shown in FIG. 8 (an example of accepting an order for fresh food by voice from the user). For a certain semantic expression, when the user makes an utterance expressing the semantic expression as content,
Input voice parameters (for example,
(1) The maximum likelihood value = 0 and the resulting semantic expression = NULL (initialization). (2) Save the input voice parameter sequence. (3) The following processing is repeated for each of the semantic expressions in the semantic expression list. (For all the semantic expressions M in the semantic expression list (the semantic expressions 1 to 3 in FIG. 2), the likelihood 1 is obtained, and the semantic expression when the highest likelihood 1 is the result is input. This is output as the meaning represented by the voice parameter sequence.) (3.1) The input voice parameter sequence and all syllable chain patterns in the sentence model (FIG. 3 syllable chain pattern numbers: 1 to 1)
4) Repeat (3.1.1) to (3.1.4) for each.

In this case, it is assumed that the last utterance from the user is a voice saying "Please display the vegetable list", and the previous utterance meaning expression is 1: (list display, *, NULL). The utterance prediction unit refers to the next utterance prediction table shown in FIG. 5 according to the input previous utterance meaning expression 1: and outputs a next utterance meaning expression appearance probability collection (0, 0.6, 0.4) to the likelihood calculation unit. . Therefore, the utterance meaning expression 1 in the likelihood calculation unit 1:
The calculation result of (list display, “vegetable” NULL) is not shown in FIG. 8 because the probability value 5 is 0 and the likelihood 1 is 0. (3.1.1) Probability value 4 = 0 (initialization). (3.1.2) The probability value 1 for the syllable chain pattern S is calculated. (3.1.3) Calculate the probability value 2 for the input speech parameter sequence.

The calculation of the probability value 1 and the probability value 2 is performed as follows. From the sentence model (see FIG. 3), conditional occurrence probabilities (0.5, 0.5, 0, 0) or (0, 0, 0.5) when syllable chain pattern numbers 1 to 4 corresponding to semantic expressions 2 and 3 occur. , 0.5) and take the probability value as 1. Next, the syllable chain patterns of syllable chain pattern numbers 1 to 4
(1: kjabetsuohitotsukudasai, ・ ・ ・, 4: chjumonowari)
Then, an acoustic feature sequence in the frequency domain of each syllable chain pattern is generated using an acoustic model (an acoustic feature model in the frequency domain of a syllable). Based on the acoustic feature sequence in the frequency domain of each generated syllable chain pattern and the input speech parameter sequence (the acoustic feature sequence in the frequency domain), syllable chain pattern numbers 1 to 3 of the speech meaning expressions 2 and 3
Probability value 2, which is the likelihood (probability) corresponding to 4: 2 (1.0 × 1
0 ^-1 , 1.0 × 10 ^-1 , 1.0 × 10 ^-4 , 1.0 × 10 ^-4 ) or (1.0 × 10 ^-4
10 ^-1 , 1.0 × 10 ^-1 , 1.0 × 10 ^-4 , 1.0 × 10 ^-4 ) are calculated.
(Calculation of probability value 2 can be performed by the above-described speech recognition method.) (3.1.4) Probability value 3 = probability value 1 × probability value 2 is calculated. (3.1.5) Probability value 4 by adding probability value 3 for each utterance expression
, That is, Σ ((probability value 1) × (probability value 2)). (Probability value 4: 1.0 × 10 ^-1 or 1.0 × the value obtained by adding over all syllable chain patterns in the sentence model
10 ^-4 . (3.2) Calculate the probability value 5 for the previous utterance meaning expressions 2 and 3. (The likelihood calculation unit transmits a request signal to the utterance prediction unit, and from the collection of semantic expression occurrence probabilities obtained as a result of the operation of the utterance prediction unit (see FIG. 4), the utterance meaning expression 2,
The next utterance meaning expression appearance probability of 3 (probability value 5): 0.6 or 0.4 is extracted. (3.3) Likelihood 1 = probability value 4 × probability value 5: 6.0 × 10 ⁻² or 4.0 × 10 ⁻⁵ for each utterance meaning expression. (3.4) Likelihood 1> Likelihood maximum value (likelihood stored before) is determined. If YES, proceed to (3.5); if NO, proceed to (3). (3.5) Maximum likelihood value = likelihood 1, as result semantic expression M (3)
Move to (4) When the above calculation is performed for all the semantic expressions in the semantic expression list, the likelihood of the resulting semantic expression is updated based on a certain standard (pre-set) (the maximum likelihood). Select and output a semantic expression for.

When the calculation result is obtained as shown in FIG.
Semantic expression number 2: Likelihood 1 (6.0 × 10 ^-2 ) of (commodity order, “cabbage”, 1) and semantic expression number 3: (order completed, NULL, NUL
L) of likelihood 1 (4.0 × 10 ⁻⁵ ) is compared, and the meaning expression with a large likelihood 1 is “(product order,“ cabbage ”,“ 1 ”)”.
Is output as a semantic expression. Further, the speech recognition device of the present invention can be configured by a computer having a CPU, a memory, and the like, a user terminal used by a user who is an access subject, and a recording medium.

The recording medium is a machine-readable recording medium such as a CD-ROM, a magnetic disk, and a semiconductor memory. The program recorded on the recording medium is read by a computer.
By controlling the operation of the computer, the components such as the voice analysis unit, the likelihood calculation unit, and the speech prediction unit are realized on the computer.

[0019]

According to the present invention, a sentence set associated with a semantic expression is used as a unified language model without dynamically creating a sentence during speech recognition, and the contents of the next utterance are set By performing statistical prediction of speech recognition and semantic analysis by predicting statistically and preventing the dispersion of probabilities due to unnecessary calculations, it is possible to accurately determine the meaning of utterance from input speech. Becomes

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration example of a speech recognition device of the present invention.

FIG. 2 is a diagram showing contents of a semantic expression list in the embodiment of the present invention.

FIG. 3 is a diagram showing a sentence model in the embodiment of the present invention.

FIG. 4 is a diagram showing a configuration example of an utterance prediction unit according to the present invention.

FIG. 5 is a diagram showing an example of the contents of a next utterance prediction table used in the present invention.

FIG. 6 is a diagram showing a likelihood calculation processing procedure according to the present invention.

FIG. 7 is a schematic diagram of a procedure for obtaining likelihood 1 for one semantic expression in the likelihood calculation unit of the present invention.

FIG. 8 is a diagram showing an example of a calculation result in a likelihood calculation unit according to the present invention.

FIG. 9 is a diagram showing a configuration example of a conventional voice recognition device.

Claims (4)

[Claims] 1. A speech recognition method for extracting a feature parameter sequence of an input speech (hereinafter referred to as an "input speech parameter sequence"), converting the input speech parameter sequence into a semantic expression, and outputting the semantic expression. The user receives the semantic expression corresponding to the user's utterance, inputs the estimated probability of occurrence of each semantic expression in the next utterance, and records the syllable standard pattern corresponding to the speech parameter sequence and the utterance sentence assumed to be uttered by the user. The likelihood of the semantic expression stored in the semantic expression list corresponding to the input speech parameter sequence is calculated by using the information of the sentence model representing the conditional occurrence probability between the syllable chain standard pattern and the corresponding semantic expression. A speech recognition method characterized by outputting a semantic expression having the highest likelihood. 2. A speech analysis unit for extracting a characteristic parameter sequence (hereinafter, referred to as an "input speech parameter sequence") of an input speech, and a semantic expression analysis unit for converting the input speech parameter sequence into a semantic expression and outputting the semantic expression. In the speech recognition device, the semantic expression analysis unit includes a semantic expression list storage unit that stores semantic expressions representing the semantic contents of an utterance sentence assumed to be uttered by the user, and an acoustic model that stores a syllable standard pattern corresponding to the voice parameter sequence. A sentence model storage unit storing a sentence model representing a conditional occurrence probability between a syllable chain standard pattern of a utterance sentence assumed to be uttered by a user and a corresponding semantic expression; An utterance prediction unit that receives a semantic expression corresponding to the utterance of the user and estimates and outputs the occurrence probability of each semantic expression in the next utterance; and a likelihood calculation unit. The likelihood calculation unit inputs the input speech parameter sequence and the probability of occurrence of the semantic expression estimated by the utterance prediction unit, and uses each information of the acoustic model and the sentence model to store the semantic expression list corresponding to the input speech parameter sequence. A speech recognition device that calculates the likelihood of the semantic expression stored in the semantic expression and outputs the semantic expression with the highest likelihood. 3. A process for extracting a characteristic parameter sequence of an input voice (hereinafter, referred to as an "input voice parameter sequence"), receiving the input voice parameter sequence and a semantic expression corresponding to the immediately preceding user's utterance, and estimating the next utterance Inputting the probability of occurrence of each semantic expression in, a condition between the acoustic model that records the syllable standard pattern corresponding to the speech parameter sequence and the syllable chain standard pattern of the utterance sentence assumed to be uttered by the user and the corresponding semantic expression The computer calculates the likelihood of the semantic expression stored in the semantic expression list corresponding to the input speech parameter sequence using each information of the sentence model representing the attached occurrence probability and outputs the most likely semantic expression to the computer. The speech recognition program to be executed. 4. A process for extracting a characteristic parameter sequence of an input voice (hereinafter referred to as an "input voice parameter sequence"), receiving the input voice parameter sequence and a semantic expression corresponding to the immediately preceding user's utterance, and estimating the next utterance Inputting the probability of occurrence of each semantic expression in, a condition between the acoustic model that records the syllable standard pattern corresponding to the speech parameter sequence and the syllable chain standard pattern of the utterance sentence assumed to be uttered by the user and the corresponding semantic expression The computer calculates the likelihood of the semantic expression stored in the semantic expression list corresponding to the input speech parameter sequence using each information of the sentence model representing the attached occurrence probability and outputs the most likely semantic expression to the computer. A computer-readable recording medium on which a speech recognition program to be executed is recorded.