JPH08314490A - Word spotting type speech recognition method and device - Google Patents

Abstract

(57) [Abstract] [Purpose] To provide a word spotting type speech recognition method and device which have a small number of source errors, high recognition accuracy, and high-speed recognition. [Structure] Using the standard pattern of vowels, the vowel candidate at each time of the input voice is obtained, and the vowel sequence in the keyword pronunciation dictionary is associated with it to detect the presence section of the keyword in the input voice (2). , Finding the likelihood of the keyword in the detected existence section by a hidden Markov model (3)
At the same time, a speech recognition result is output from the detected keyword existence section and the obtained likelihood (4).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition method and apparatus, and more particularly to a word spotting type voice recognition method and apparatus for detecting a keyword in an input voice to perform voice recognition.

[0002]

2. Description of the Related Art As a method of using a conventional word spotting type voice recognition device, for example, the following methods (a) and (b) have been proposed. (A) Method using only Hidden Markov Model (See Akihiro Imamura: "Spotting of Telephone Voice by HMM (Hidden Markov Model)", IEICE Technical Report SP90-18, 1990). (B) A method in which word preselection by a phoneme recognition network element is combined with word determination by DP matching (for example, "speech recognition by a probabilistic model", published by The Institute of Electronics, Information and Communication Engineers, Sei Nakagawa, page 18, 1988) (Ogura Hiromi: “Word preselection and large vocabulary word speech recognition using phoneme posterior probability time series,” IEICE Tech. SP94-87, 1995).

[0003]

In the above-mentioned conventional method (a), it is assumed that an arbitrary frame to an arbitrary frame of the input speech is a keyword existence section, and the likelihood is calculated only by the hidden Markov model. Therefore, there is a problem that the amount of calculation is large and that a keyword is often detected (a spelling error) in a place where it does not actually occur.

Further, with the conventional method (b), it is difficult to realize recognition of all phonemes including consonants with high accuracy, and the accuracy of likelihood calculation of keywords by DP matching is also particularly high. There was a problem that it was inferior to Hidden Markov Model when assumed. SUMMARY OF THE INVENTION Therefore, an object of the present invention is to eliminate the drawbacks of the conventional techniques of the related art, to provide a word spotting type speech recognition method and device which have few source errors, high recognition accuracy and high-speed recognition. .

[0005]

In order to achieve this object, the word spotting type speech recognition method of the present invention finds a vowel candidate at each time of an input voice using a standard pattern of vowels and stores it in a keyword pronunciation dictionary. Corresponding to the vowel sequence, the presence section of the keyword in the input speech is detected, the likelihood of the keyword in the detected presence section is obtained by the hidden Markov model, and the presence section of the detected keyword is obtained. The speech recognition result is output from the likelihood.

The word spotting type speech recognition apparatus of the present invention further comprises an acoustic analysis section for performing acoustic analysis of input speech,
The vowel recognition of the input voice is performed based on the analysis result of the analysis unit and the keyword pronunciation dictionary provided in advance,
An existence section detection unit that detects an existence section of each keyword in the input voice, and a likelihood calculation unit that obtains the likelihood of each keyword included in the input voice based on the analysis result, the keyword pronunciation dictionary, and the detected keyword existence section. When,
The present invention is characterized by further comprising a recognition result output unit that outputs a keyword recognition result based on the presence interval of each keyword and the obtained likelihood.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described in detail below with reference to the accompanying drawings. Referring to FIG. 1 showing an embodiment of an apparatus used for implementing a word spotting type speech recognition method according to the present invention, the speech recognition apparatus of the present invention comprises an acoustic analysis unit 1 for performing acoustic analysis of input speech. The presence section detection unit 2 for recognizing vowels based on the analysis result and the keyword voice dictionary and detecting the presence section of each keyword in the input voice, and for each keyword based on the analysis result, the keyword voice dictionary, and the keyword presence section. It is composed of a likelihood calculation unit that obtains a likelihood and a recognition result output unit that outputs a keyword recognition result based on the existence interval and likelihood of each keyword.

As shown in FIG. 2, the acoustic analysis unit 1 A / D-converts an input voice by an A / D converter 11, and a framer 12 sets a width of about 20 ms as one frame.
The calculation unit 1 calculates values such as the average power of each frame and the LPC (linear prediction) cepstrum coefficient (coefficient representing frequency distribution).
Calculate by 3.

As shown in FIG. 3, the existing section detecting unit 2 previously stores standard values (vowel standard patterns) such as cepstral coefficients of vowels (a, i, u, e, o, and n) as vowel standard patterns. It is saved in the memory unit 21. The vowel standard pattern and the analysis result of the input voice are compared by the comparator 22, and the vowel candidate determining unit 23 determines which vowel each frame of the input voice corresponds to. However, instead of uniquely determining one vowel for each frame, the proximity of the standard pattern of each vowel and the input voice is converted into a likelihood to obtain a plurality of vowel candidates for each frame.

Next, it is checked based on the keyword pronunciation dictionary whether or not the vowel sequence of the keyword to be recognized is included in the input voice. This is also performed by the vowel candidate determination unit 23. The keyword pronunciation dictionary is a description in which the keywords to be recognized are mixed with kanji and kana and their pronunciations are written in Roman letters. For example, the pronunciation "s, u, p, oo, ts, u" for the keyword "sports" and the pronunciation "s, u, m, oo" for the keyword "sumo" are described in advance. I'll do it. Keyword vowel sequence "u,
When the input voice includes "oo, u", "u, oo", etc., it outputs which section of the input voice is from which frame to which keyword, and the existence section information of each keyword is output.

When checking whether a vowel sequence of a keyword is included in an input voice, for example, when checking whether a vowel sequence “u, oo, u” of “sports” is included in an input voice, “u, Even when a vowel is inserted such as "oo, a, u", it is also defined as a keyword existence section. Of course, a plurality of existing sections can be output for one keyword. The above operation is performed for all the keywords in the keyword pronunciation dictionary.

As shown in FIG. 4, the likelihood calculating unit 3 has a hidden Markov model for each phoneme of a vowel (a, i, u, e, o) and a consonant (p, t, k, ts) in advance. Save in memory (ROM). Then, these are serially connected 32 according to pronunciations in the keyword pronunciation dictionary, and a hidden Markov model corresponding to a keyword such as "sports" or "sumo" is created in a block 33 (RAM). The existence section of each keyword is received from the existence section detection unit 2, and the likelihood of the input speech with respect to the section is obtained by the likelihood calculator 34 by the hidden Markov model corresponding to each keyword.

As shown in FIG. 5, the recognition result output unit 4 receives the presence interval and likelihood of each keyword in the likelihood memory unit 41, and the likelihood threshold from the preset threshold memory unit 42 and the comparator 43. When the comparison is made and there is a keyword exceeding the threshold value, it is determined that the keyword is uttered, and the keyword is output as a recognition result.

A recognition example will be described with reference to FIG. For example, when a voice such as "Are there any sports programs?" Is input, the presence section detection unit 2 first recognizes vowels, and then "a, oo, u, oo, u, a, N, u, Vowel sequences such as "i, a, i, a, u, a" are obtained. Next, "Sports (s, u, p, oo, t
s, u) '', `` ski (s, u, k, ii) '', and `` sumo (s, u, m, oo) '' are detected as matching the vowel sequence of the input voice. This section is passed to the likelihood calculation unit 3 as a keyword existing section. Likelihood calculation unit 3 uses “sports (s, u, p, oo, ts, u)”, “skis (s, u, k, ii)”, and “sumo”.
For (s, u, m, oo) ”, the likelihood is calculated by the hidden Markov model. The recognition result output unit 4 shows the largest likelihood,
Likelihood to exceed a predetermined threshold such as -13.0-
“Sports” indicating 12.0 is output as the keyword recognition result.

The effectiveness of the present invention is verified by recognition experiments. First, the vowel standard pattern and hidden Markov model were learned by a commercially available female voice database, and speaker adaptation was further performed with 50 sentences of two female speakers. The LPC cepstrum coefficient was used as a feature amount for the vowel standard pattern and the hidden Markov model. Twenty word keywords were registered in the keyword pronunciation dictionary. The two female speakers mentioned above
Forty-eight sentences including one keyword in the sentence were uttered and recognized by the recognition device of the present invention. As a result, a recognition rate of 94% was obtained (it was 30% in the conventional method (a)), and there was no springing error, and the processing was completed in almost real time.

Although the present invention has been described with reference to one embodiment, it is obvious to those skilled in the art that the present invention is not limited to this embodiment and various modifications and changes can be made within the scope of the present invention. Will. For example, the likelihood 24 shown in FIG. 3 may be obtained by comparing the vowel pattern of the input speech as the analysis result with the vowel standard pattern to obtain the distance difference between the two. In addition to the hardware configuration described above, the apparatus for implementing the method of the present invention may be processed by programmed computer software.

[0017]

According to the method and apparatus of the present invention, the likelihood of the keyword is calculated using the hidden Markov model after the existence section of the keyword in the input speech is calculated using the standard pattern of vowels. Provided are a voice recognition method and a device capable of high-speed recognition with less generation error and higher recognition accuracy than a spotting type voice recognition device.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of an apparatus for carrying out the method of the present invention.

FIG. 2 is a diagram showing a configuration of an acoustic analysis unit shown in FIG.

FIG. 3 is a diagram showing a configuration of an existing section detection unit shown in FIG.

FIG. 4 is a diagram showing the configuration of a likelihood calculation unit shown in FIG.

FIG. 5 is a diagram showing a configuration of an illustration recognition result output unit in FIG.

FIG. 6 is a diagram for explaining a recognition example.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Acoustic analysis unit 2 Presence section detection unit 3 Likelihood calculation unit 4 Recognition result output unit 11 A / D converter 12 Framer 13 Power LPC cepstrum coefficient calculation unit 21 Vowel standard pattern memory unit 22 Comparator 23 Vowel candidate determination unit 24 Likelihood 31 Hidden Markov Model Memory Unit of Phoneme Unit 32 Concatenation 33 Hidden Markov Model of Keyword Unit 34 Likelihood Calculator 41 Likelihood Memory Unit 42 Threshold Memory Unit 43 Comparator

Claims (2)

[Claims] 1. A vowel candidate at each time of an input voice is obtained using a standard pattern of vowels, the vowel sequence in a keyword pronunciation dictionary is associated with the vowel candidate, and a keyword existing section in the input voice is detected. A word spotting type speech characterized in that a likelihood of a keyword in the detected existence section is obtained by a hidden Markov model, and a speech recognition result is output from the detected existence section of the keyword and the obtained likelihood. Recognition method. 2. The presence of each keyword in the input voice, which recognizes the vowels of the input voice based on the acoustic analysis unit that performs the acoustic analysis of the input voice, the analysis result of the analysis unit, and the keyword pronunciation dictionary provided in advance. An existing section detection unit that detects a section, a likelihood calculation section that obtains the likelihood of each keyword included in the input voice based on the analysis result, the keyword pronunciation dictionary, and the detected keyword existing section, and the existing section of each keyword And a recognition result output unit that outputs a keyword recognition result based on the obtained likelihood, and a word spotting type speech recognition device.