JPS6312000A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a speech recognition device for non-specific speakers that derives recognition candidate speech.

Conventional technology In general, in speech recognition devices for non-specific speakers, a large number of speech patterns from a large number of people are divided into groups using a clustering method, their representative patterns are registered as standard patterns, and the input speech patterns are combined with the input speech patterns stored in a dictionary. A method is used in which the degree of similarity is calculated between all the standard patterns, and then the standard pattern with the maximum degree of similarity is selected as the recognition candidate speech. In order to calculate the similarity between two speech patterns, dynamic programming (
Pattern matching (hereinafter referred to as DP matching) is used in which the time axes of two patterns are expanded or contracted non-linearly using a dynamic programming method. In particular, word speech recognition devices achieve high recognition rates using this DP matching method. (For example, [Dynamic Programming Optimization for Spoken Word Recording, 7 J (H, 5 akoa
and S, C:hiba.

” Dynamic programming Opt.
imization rosporkan word
recognition”, ENGEEE trans
.

personcoustic, 5peech, Signal P
rocessing, Vol.

Mu5SP-27pp, 336-349.1979))
Problems to be Solved by the Invention However, with the above-mentioned speech recognition device, when a pattern with a missing beginning or end of a word is input due to individual differences in the way speakers pronounce their utterances, errors in speech segment detection, etc. This method has the problem that pattern matching is performed with patterns that have no omissions, resulting in a low degree of similarity and an increased likelihood of misrecognition.

For example, if we say ``rFUKUOKA (Fukuoka)'', the FH part at the beginning of the word may be voiced or devoiced depending on the way of pronouncing9 between individuals.0 If it is devoiced, the energy value of the FU part becomes very small, and in a speech recognition device that mainly detects voice sections based on the energy value series of the voice, rKUOK where the FU part is mistakenly omitted.
There is a high possibility that only the section "A" will be detected as a voice section. Therefore, even if pattern matching is performed with the standard pattern rFUKUOKA, the degree of similarity will be low and misrecognition will likely occur. As described above, the problem with conventional speech recognition devices is how to prevent the recognition rate from decreasing when a speech section is incorrectly detected.

In view of the above-mentioned problems, the present invention has been developed to improve the sound of the beginning part of a word depending on the way it is uttered.
To provide a speech recognition device that can accurately recognize a pattern in which there is a possibility of missing a word ending even if a speech section is incorrectly detected.

Means for Solving the Problems In order to achieve the above object, the speech recognition device of the present invention includes a speech section detecting means for detecting a speech section from the energy series of input speech, and a speech section detecting means for detecting a speech section from the energy series of input speech, and a speech section detecting means for detecting a speech section from a large number of speech patterns of a large number of people. standard pattern determining means for selecting a plurality of patterns for each voice to be recognized and determining them as standard patterns; standard pattern management means for managing the stored address and pattern length of the standard patterns; and nine individual utterance methods. A partial pattern management means that manages a pattern in which there is a possibility that the initial part of a word or a final part of a word is missing due to the difference as a part of a standard pattern without any missing parts; The apparatus is equipped with a pattern matching means that performs pattern matching between each pattern and selects the pattern with the maximum degree of similarity as a recognition candidate speech.

Effects of the Invention With the above-described configuration, the present invention manages patterns in which there is a possibility of missing the beginning and end of a word in advance as a part of a standard pattern without missing parts, and creates a standard pattern without missing parts. By performing pattern matching between the partial pattern of the missing representative pattern and the input speech and deriving recognition candidate speech, it is possible to detect speech sections incorrectly for patterns that are difficult to detect at the beginning or end of two words. can also be recognized with high accuracy. Furthermore, by managing a pattern with a missing part as a part of a standard pattern without any missing part, it is possible to prevent the memory capacity of the standard pattern from increasing.

Embodiment Hereinafter, a speech recognition device according to an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention. In FIG. 1, reference numeral 1 denotes a voice input section, into which a speaker's voice is input through a microphone or the like. 2 is a speech analysis means that extracts a time series and an energy series of feature vectors from the input speech signal. 3 is a voice section detecting means that detects a voice section portion from the energy sequence of the voice.

Reference numeral 4 denotes a standard pattern determining means that analyzes a large number of voice patterns from a large number of people and determines their representative patterns as standard patterns. Reference numeral 6 denotes a standard pattern management means for managing the memory location and pattern length of each standard pattern; 6 a part for managing patterns with missing two word beginnings and tails as part of the standard pattern managed by the standard pattern management means 6; Pattern management means, 7
8 is a pattern matching means that performs pattern matching between the input pattern and each standard pattern and each partial pattern, and 8 is a recognition device that notifies the speaker of the recognition candidate speech derived from the result of the pattern matching means 7 through speech synthesis or the like. This is the result output section.

FIG. 2 is a circuit diagram showing the configuration of this embodiment, which includes the above-mentioned voice section detection means 3, standard pattern management means 6.degree. partial pattern management means 6. This figure shows a configuration in which the pattern matching means 7 is realized by a microcomputer 23. In FIG. 2, reference numeral 11 is a microphone that inputs audio, and 12 is an analog/digital converter (hereinafter referred to as A) that converts the audio signal input from the microphone 11 from analog to digital.
/D converter. ), 13 is a speech analysis unit, 14 is a speech interval detection unit, 15 is an input pattern memory that stores the time series of feature vectors of input speech, and 17 is a possibility that one word beginning or end part is missing in the standard pattern. Regarding a certain pattern,
A standard pattern partial pattern management table for managing missing patterns as partial patterns of the standard pattern; 18 a standard pattern management table for managing each standard pattern determined by the standard pattern determining means 6; 19 a standard pattern management table for managing all standard patterns. 20 is a recognition result determination section; 21 is a speech synthesis section that synthesizes the speech of the obtained recognition candidate speech; 22 is a speech synthesis section 21
This is a speaker that outputs the voice synthesizer obtained by.

FIG. 3 is a main part flowchart for explaining the operation of the microcomputer of this embodiment.

The operation of this embodiment with the above configuration will be explained in detail with reference to the flowchart of FIG.

First, in step 31, audio is input from the microphone 11, the audio signal is converted from analog to digital by the human/D converter 12, and then the audio analysis unit 13 converts the audio pattern into a feature vector (for example, a 10-dimensional linear prediction coefficient). Find the time series and energy series of. In step 32, from the energy series obtained by the speech analysis unit 13, the sections in which the energy value exceeds the threshold value exceed T8fc for a certain period of time, and furthermore, the sections in which the energy value exceeds the threshold value exceed T8fc for a certain period of time T1. Threshold people above T2. When there is an interval below T for a certain period of time.

The section exceeding the above is detected as a speech section, and in step 33, the time series of the feature vector is stored in the input pattern memory 16.

Note that the standard pattern determining means 4 determines in advance a plurality of representative patterns from among the many voice patterns of many people for each recognition target voice, and stores them in the standard pattern memory 1.
9 memorizes those patterns. Further, the standard pattern management table 18 stores the addresses and pattern lengths for managing each pattern in the standard pattern memory 19, and the standard pattern partial pattern management table 17 stores the beginning of the word of the standard pattern. In order to check in advance a pattern that may have a missing two-word tail, and to manage the pattern when the missing part occurs as a partial pattern of a standard pattern without missing parts, the address on the standard pattern memory 19 and its pattern are stored. 0 that stores the length, that is, the standard pattern memory 19
only stores the time series of feature vectors of the standard pattern as a representative pattern with no omissions, and when a missing partial pattern at the beginning or end of a word is needed, partial pattern management of the standard pattern is performed. It is sufficient to extract only the partial pattern portion from the standard pattern memory 19 according to the table 17.

In step 34, the first pattern on the standard pattern memory 19 is loaded into the memory of the DP matching unit 16 according to the standard pattern management table 18, and then in step 35, the input pattern stored in the input pattern memory 16 and the input pattern are loaded in step 34. DP matching is performed between the standard pattern and the standard pattern. In step 36, standard pattern management table 1
8, it is checked whether all the standard patterns and the processing of steps 34 and 35 have been completed, and if not, the process returns to step 34 and the same processing is continued.

If the conditions in step 36 are satisfied, then in step 37 the first partial pattern is loaded from the standard pattern memory 19 onto the memory of the DP matching section 16 according to the partial pattern management table 17, and in step 38 DP matching is executed. Thereafter, in step 39, it is checked in accordance with the standard pattern partial pattern management table 17 whether all partial patterns and the processes of steps 37 and 38 have been completed, and if not, the process returns to step 37.

When the DP matching with all standard patterns and partial patterns is completed, the process proceeds to step 4o, where the recognition result determination unit 2
At step o, the pattern that gives the maximum value among the degrees of similarity with each standard pattern and partial pattern obtained by the DP matching unit 16 is determined as a recognition candidate speech. Further, in step 41, the speech synthesis section 21 is started to synthesize the recognition candidate speech obtained by the recognition result determination section 2o, and outputs the synthesized speech to the speaker 22 to notify the speaker of the recognition candidate speech.

In this embodiment, the standard pattern management table and the partial pattern management table are provided separately, but if the partial pattern management table is considered as part of the standard pattern management table, the same processing can be performed with a single management table. It can be carried out.

As described above, according to this embodiment, the standard pattern management means for managing standard patterns, and the standard pattern management means for managing standard patterns, and the standard pattern that does not have any deletions, are used to convert the pattern when the beginning or end of a word is likely to be deleted. Even if a word beginning or two word endings are mistakenly detected, correct recognition can be performed by pattern matching with the partial pattern.

Furthermore, regarding the unstable standard patterns of the beginning and end of words, the missing patterns can be managed with a single representative pattern that is free of omissions, so there is no need to increase the number of templates, and memory can be used effectively.

Effects of the Invention As described above, the present invention selects a plurality of representative patterns from a large number of speech patterns of a large number of people for each speech to be recognized,
A standard pattern determining means for determining a standard pattern, a standard pattern management means for managing the memory address and pattern length of each standard pattern, and a pattern that may be missing two word beginnings and tails among the standard patterns. It has a partial pattern management means that manages the address and pattern length of a pattern when a missing part occurs as part of a standard pattern without missing parts in one standard pattern. By performing pattern matching between the two and selecting the pattern with the highest degree of similarity as the recognition candidate speech, it is possible to manage partial patterns even when a pattern in which the beginning or end of a word is accidentally omitted during speech section detection is detected. By performing pattern matching with the partial patterns managed by the above method, it is possible to provide a speech recognition device that can perform recognition with high accuracy.

In addition, by managing missing patterns as a single representative pattern, using a standard pattern with no missing parts as a representative pattern,
It is possible to provide a speech recognition device that can correctly recognize even when a speech section is detected incorrectly without increasing the number of templates.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a speech recognition device according to an embodiment of the present invention, FIG. 2 is a circuit block diagram showing the configuration of the same device, and FIG. 3 is a flowchart of main parts for explaining the operation of the same device. It is. 2... Voice analysis means, 3... Voice section detection means, 4... Standard pattern determining means, 6...
... Standard pattern management means, 6 ... Partial pattern management means, 7 ... Pattern matching means, 1
1... Microbon, 15... Input pattern memory, 17 Old... Partial pattern management-y--7"
18...Standard pattern management table, 19.
...Standard pattern memory, 22... River speaker,
23...Microcomputer.

Claims (1)

[Claims] speech analysis means for extracting a time series of feature vectors including energy sequences from input speech; speech section detection means for detecting speech sections from the energy sequences obtained by the speech analysis means; standard pattern determining means for selecting representative patterns and determining a plurality of them as standard patterns for each voice to be recognized; and an address in a memory where each standard pattern determined by the standard pattern determining means is stored; Regarding the standard pattern management means for managing pattern length and the standard pattern of recognition target speech where the beginning or end of a speech pattern may be missing depending on the way of speaking or individual differences, a pattern with no missing parts is used as a representative pattern. , a partial pattern management means for managing the memory address and pattern length where the standard pattern of the missing pattern is stored as a part of the representative pattern; each standard pattern managed by the standard pattern management means; The present invention is characterized by comprising a pattern matching means that performs pattern matching between each partial pattern of the standard pattern managed by the partial pattern management means and an input speech pattern, and sets a pattern with a maximum degree of similarity as a recognition candidate speech. voice recognition device.