JP2005189294A - Speech recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition device capable of recognizing speech without removing noise even if the noise such as breathing or lip noise is included in a speech waveform. <P>SOLUTION: A speech analyzer 12 analyzes speech spoken by a user and outputs a speech waveform signal, a syllable segmenting unit 14 cuts a monosyllabic part out of the speech waveform signal, and a syllable recognition unit 2 recognizes the cut out monosyllabic part. Then the recognized character string and a character string obtained by deleting at least one character from the recognized character string are matched with a recognition dictionary. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech recognition device, and more particularly, to a speech recognition device that recognizes speech based on speech waveforms that are sounded separated into single syllables.

2. Description of the Related Art Conventionally, a speech recognition apparatus that recognizes speech based on a speech waveform uttered by dividing a word into syllables is known. In this speech recognition device, misrecognition occurs when breath or lip noise is detected. Therefore, the syllable is determined from the syllable candidates by threshold comparison processing, and the breath or lip noise other than the necessary syllable is removed as noise. (For example, Patent Document 1), breathing sounds, lip noise, and the like are removed using a voiced hidden Markov model and a silent hidden Markov model (for example, Patent Document 2).
Japanese Patent Laid-Open No. 7-261799 JP 11-288293 A

  However, in the above-described conventional technique, since the breath or lip noise is recognized based on the features such as the breath or lip noise in the voice section detection stage in the signal processing and is removed as a sound other than the voice, it is complicated. There is a problem in that signal processing is necessary and voice may be mistakenly recognized as noise and dropped.

  The present invention has been made to solve the above-mentioned problems, and a speech recognition apparatus capable of performing speech recognition without removing these noises even if the speech waveform contains noises such as breath and lip noise. The purpose is to provide.

  In order to achieve the above object, the present invention provides a voice analysis means for analyzing a voice uttered by a user and outputting a voice waveform signal, a syllable extraction means for cutting out a syllable part from the voice waveform signal, and an acoustic model Recognition means for performing speech recognition of each syllable part extracted using, outputting a character string, a character string recognized by the recognition means, and a character string from which at least one character has been deleted from the character string And matching means for matching with the recognition dictionary.

  The speech analysis means of the present invention analyzes speech uttered by the user and outputs a speech waveform signal, and the syllable extraction means extracts a syllable portion from the speech waveform signal. The recognition means performs speech recognition of each syllable portion extracted using the acoustic model and outputs a character string, and the matching means obtains a character string recognized by the recognition means and at least one character from the character string. The deleted character string is matched with the recognition dictionary. And a matching means outputs the character string matched with the dictionary as a recognition candidate.

  If the character string does not match the recognition dictionary even if one character is deleted, matching is performed by increasing the number of characters to be deleted until the character string from which the character has been deleted matches the recognition dictionary. , Recognition candidates can be output.

  As described above, according to the present invention, since the character string obtained by deleting at least one character from the character string recognized by the recognition unit is matched with the recognition dictionary, breath and lip noise are included in the speech waveform. Even if noises such as the above are included, an effect that voice recognition can be performed without removing these noises can be obtained.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  As shown in FIG. 1, the present embodiment includes a microphone 10 that detects a voice uttered by a user dividing a word into syllables, and voice power information obtained by analyzing a voice detected by the microphone 10. A voice analyzer 12 for outputting a voice waveform signal is provided.

  The speech analyzer 12 is connected to a syllable extractor 14 that extracts a speech segment (single syllable portion) from the input speech waveform signal and outputs a syllable string signal.

  A syllable recognizer 16 to which a storage device 18 is connected is connected to the syllable extractor 14. The storage device 18 includes “a”, “i”, “u”, “e”, “o”, “ka”, “ki”, “ku”, “ke”, “ko”, “sa”, , Etc. 50, which are represented in the Japanese syllabary diagram such as “ga”, “gi”, “gu”,..., Etc., “pa”, “pi”, “pu”,. A total of 110 basic Japanese syllable acoustic models (isolated syllable models) including semi-voiced sounds such as “Kya”, “kyu”, “kyo”,... .

  The syllable recognizer 16 is connected to a character string matching unit 20 to which a storage device 22 that stores a large number of words as a recognition dictionary is connected.

  Next, the operation of the speech recognition apparatus according to this embodiment will be described. The voice uttered by the user is detected by the microphone 10, the voice detected by the microphone 10 is input to the voice analyzer 12, and the voice analyzer 12 analyzes the single syllable utterance uttered by the user divided into syllables. Then, a feature quantity such as voice power information is extracted and a voice waveform signal W shown in FIG. 2 is output. FIG. 2 shows a speech waveform signal when “I, Tsu, Shi, N, Sha, (nasal breath)” is uttered, and nasal breath is generated at the end of the utterance. Included.

  The speech waveform signal output from the speech analyzer 12 is input to the syllable extractor 14. The syllable extractor 14 compares the input speech waveform signal with the power threshold value, and extracts a portion where the power is greater than the threshold value as a speech section, and outputs the syllable string signal S shown in FIG. Since the power (amplitude) of the speech waveform signal is small between syllables, the speech section can be cut out by setting a threshold larger than the portion where the power is small.

  At this time, if the power threshold is set high, breath and lip noise are often not detected, but syllables uttered by a small voice are also lost at the same time, so the threshold should be set low. I do not get. In the example of FIG. 2, the nasal breath uttered at the end of the utterance is also erroneously cut out as a syllable section.

  The syllable recognizer 16 uses the single syllable acoustic model stored in the storage device 18 to perform speech recognition of a single syllable portion extracted as a syllable section. In the example of FIG. 2, if anything other than nasal breath is correctly recognized, it is recognized as “I, tsu, shi, n, sha, (nasal breath)”. This (nasal breath) portion is recognized as, for example, “fu”. For this reason, in a normal process, it will not be correctly recognized in the subsequent character string matching process.

  Therefore, in the present embodiment, in the character string matching unit 20, not only the recognition result by the syllable recognition unit 16 but also n characters (for example, 1 or 2) from the character string of the recognition result by the syllable recognition unit 16 are used. ) Matching is also performed on the deleted character string, and a matching recognition candidate is output. In this case, matching is performed by sequentially increasing the number of characters to be deleted until matching with the dictionary.

  In the example of FIG. 2, since the character string is recognized as “I, Tsu, Shi, N, Sha, Fu”, first delete one syllable (one character), and “Tsu, Shin, N, Sha, Fu” ”,“ I ・ Sh ・ n ・ Sha ・ fu ”,‘ I ・ tsu ・ n ・ sha ・ fu 」,‘ I ・ tsu ・ shi ・ sha ・ fu 」,` `I ・ tsu ・ shi ・ sha ・ fu '', String matching is performed for the six types “I, Tsu, Shi, N, and Sha”. In this case, since the sixth “Itsutsu, Shin, Sha” matches the dictionary, the matched “Itsu, Shi, Shin, Sha” is output as a recognition candidate.

  If a recognition candidate is not obtained even if one syllable is deleted, one syllable, that is, two syllables, is deleted and matching processing similar to the above is performed. The deletion of syllables and the matching process after deletion are repeated until a recognition candidate is obtained.

  When noise such as breath and lip noise is not included, matching with the recognition dictionary is possible without deleting the syllable from the character string of the recognition result by the syllable recognizer 16.

  Furthermore, if it is recognized as “Mi / ka / n / fu” (the last “fu” is a nose breath), deleting one syllable will result in “ka / n / fu”, “mi / n / fu”, “ "Mi-ka-fu", "mi-ka-n", and when these are matched with a dictionary, two or more characters "ka-n-fu" and "mi-ka-n" It is also possible that the columns match. In such a case, “mi / ka / n” may be selected as a recognition candidate in consideration of the appearance frequency of words in the sentence.

It is a block diagram which shows embodiment of this invention. It is a wave form diagram which shows the example of a syllable string signal when a speech waveform signal and a speech waveform signal are cut out for every syllable part.

Explanation of symbols

10 microphone 12 speech analyzer 16 syllable recognizer 20 character string matcher

Claims (2)

A voice analysis means for analyzing a voice uttered by a user and outputting a voice waveform signal;
Syllable cutting means for cutting out a syllable portion from the speech waveform signal;
A recognition means for performing speech recognition of each syllable portion extracted using an acoustic model and outputting a character string;
A matching means for performing matching with a recognition dictionary for a character string recognized by the recognition means and a character string obtained by deleting at least one character from the character string;
A speech recognition device.   The speech recognition apparatus according to claim 1, wherein the matching unit performs matching by increasing the number of characters to be deleted until the character string matches the recognition dictionary.

Images