JPS6344700A - Word detection system - Google Patents

Abstract

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

Description

[Detailed Description of the Invention] (Industrial Application Field) The present invention is used in speech recognition devices, speech input devices, etc., and detects words included in input speech and the position of the word in the speech. Concerning word detection methods.

(Prior Art) A method for detecting a single character 5 and its position in an input voice in a voice recognition device, voice input device, etc. 1. Each category and its position information in the input speech are extracted from the input speech, which is a series of categories such as syllables, phonemes, phoneme classes, etc., and the categorization 1 of the word has a category series created from the extracted categories. ], there is a method that uses the word and the position of the first category string in the input speech as the detection result.

General 1: The above-mentioned categories have short durations and similar categories exist, so it is difficult to extract the categories from the input speech completely without error. For this reason, conventionally, multiple category candidates are extracted for each category section in the input audio, and a partial category candidate sequence is generated using the category candidates from the end of the input audio to the neck. By repeating the process of comparing a word with the category tag q, a category candidate string corresponding to the word was found. The details of this method are described in, for example, Document 11, Japanese Patent Application No. 58/214544, "Bang Recognition Apparatus", and will therefore be omitted here.

In addition, during the stage of category extraction from input speech, a certain category may disappear due to reasons such as lax pronunciation or articulatory combination between adjacent categories (for example, syllables), or its existence may not be detected. The categories before and after the category may be extracted as adjacent categories.

This phenomenon will hereinafter be referred to as category dropout.

In order to deal with this, conventional methods have been to investigate in advance the arrangement of categories in which category dropouts occur, and based on the results, to determine which category dropouts occur with a relatively high frequency. A category 9 correction rule for converting a category string into a category string in which dropped categories are restored is prepared. This rule
By applying this method to a category candidate string when detecting a word, it is possible to restore a dropped category even if the category is dropped with a relatively high frequency. For more information on this method, see
For example, Reference 2r Namiki, Hamada, Nakatsu, “Japanese Input Method Using Speech Recognition,” Theory of IEICE Vol. 1. J67D
No. 4, 1984.4 J, so it will be omitted here.

(Problem to be Solved by the Invention) In the above conventional method, a category candidate string is generated using category candidates extracted from input speech, and then a match is made with a word category string. This results in a large number of wasteful category candidate sequences being generated, which requires a large amount of calculation.

In addition, even if the section of the word to be detected occupies only a part of the input speech, conventionally, all category candidates are equally matched against the categories in the word from the end of the input speech, including the section where the word does not exist. must be done,
This requires unnecessary calculation time and requires a long time until I-i is detected.

Furthermore, the above category string correction rule
This applies not only to category candidate sequences in which g occurs, but also to category candidate sequences in which g does not occur. Furthermore, in many cases, a plurality of correction rules can be individually applied to - category candidate columns. For this reason, many category candidates are generated from one category candidate sequence, and they must be detected! Many candidate category sequences must be examined until a category candidate sequence corresponding to the X word is found. Moreover, most of the category candidate strings do not match the category string of the word to be detected and are ultimately wasted.

Further, the categories that can be restored using the correction rules are limited to those due to category omissions that occur relatively frequently, and it is impossible to restore categories that occur relatively infrequently. In order to increase the number of omissions that can be restored, the number of correction rules must be increased, and as a result, the number of category candidate sequences that can be generated increases.

For example, when I tried to extract syllable candidates from a voice that had been pronounced as ``Onseininshikitwa (What is speech recognition?),'' I found that the syllable ``shi'' was pronounced!!The waiting time was short, and the syllable ``
Because the syllable boundaries between ``shi'' and the syllables ``n'' and ``ki'' before and after it were close, the presence of the syllable ``shi'' could not be detected, and the syllables ``n'' and ``ki'' before and after it could not be detected. Assume that syllable candidates are extracted at adjacent positions.As a result, even if correct syllable candidates are obtained for all syllables other than the syllable "shi", the generated syllable candidate string is The syllable ``shi'' which becomes towa and is dropped must be restored using correction rules. However, this kind of syllable omission is a relatively rare type, and it seems that there are few rules in place to correct this omission. Also, even if this correction rule is prepared, there are still other errors, such as “nin” → “niin”, “i” → “ii”.
There are many correction rules that are prepared, and by applying these, “Onsei Inin Kitwa”,
Useless syllable candidate strings such as "once in intowa" and "once ini intowa" are also generated.

It is an object of the present invention to avoid generating unnecessary category candidate sequences, and to avoid generating unnecessary category candidate sequences when the range of words to be detected is a small part of the entire input speech, or when the number of words to be detected in the input speech is Even if a category is dropped, it is possible to efficiently detect words and the position of -1 from input speech! The object of the present invention is to provide an X word detection method.

(Means for Solving the Problems) In order to solve the above-mentioned problems and achieve the above objects, the present invention provides means for solving the above-mentioned problems and achieving the above objects. Using multiple category candidates and their positional information, word category [
By generating a category candidate sequence corresponding to 1,
This is a word detection method that detects words and their appearance positions in input speech. Each of a plurality of category candidates obtained from input speech is classified and written by its category name, and the order of categories in a word is calculated. Accordingly, a category candidate corresponding to each category is selected from among the category candidates stored and classified with the same name as that category, and the first and last categories in the sequence of three consecutive categories in the word are When corresponding to each category candidate in a sequence of two consecutive category candidates in the input audio, a sequence of category candidates is generated by making the sequence of three categories correspond to the sequence of two category candidates. It is characterized by carrying out the following.

(Operation) The method of the present invention uses only category candidates with the same name as the category included in the word to be detected from among the category candidates extracted from the input speech, and responds by tracing the arrangement of categories in the word. Generate category candidate taro. As a result, only category candidate sequences corresponding to word category sequences or subsequences thereof are generated, making it possible to avoid generating unnecessary category sequences.

In addition, in order to generate a category candidate string from the category candidates corresponding to the categories in words among the category candidates in the input speech, it is necessary to Even if the word exists, and no matter where the section is located in the input speech, it is possible to quickly detect the word.

If a certain category of words in the input audio is dropped,
Category candidates for two categories adjacent to each other before and after that category are adjacent to each other in the input speech.

In other words, the category string in a word is C+-I CI CI
ol, and if category C1 drops out, C++I becomes CI
-1, each category candidate is extracted.

Therefore, when generating a category candidate string corresponding to the word while tracing the arrangement of categories in the word to be detected, the category candidate corresponding to CI+H above is Cl−
If the category candidate corresponding to 1 follows in the input speech, the arrangement of the two category candidates is made to correspond to the category string C+-+ C, C, . . . in the word. As a result, even if category C5 is dropped, it is possible to maintain correct correspondence between the category candidate string and the category string.

Furthermore, since only the category candidate string corresponding to the word category string can be generated, it is possible to avoid wasteful generation of category candidate strings.

(Examples) Hereinafter, the present invention will be described in more detail according to examples with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In this example, it is assumed that Japanese audio is input.
Use syllables as categories.

The syllable extraction unit 1 detects syllable candidates in input speech and stores the candidates in the syllable candidate storage unit 2. As an example, assume that the voice "Onseininshikitoha" (what is the voice T2m) is input. In this case, syllable candidates such as those shown in the second section are extracted as a result of syllable recognition. In FIG. 2, arrow lines indicate sections of each syllable candidate, and a plurality of syllable candidates are extracted for each section. These syllable candidates are classified by syllable person and stored in the syllable candidate storage section 2. As a result, the contents of the syllable candidate storage section 2 become as shown in FIG. In this figure, each syllable candidate is expressed in the format of "syllable base/start time: end time".

The word storage unit 3 stores syllable strings of words to be detected. After one word is extracted into the word buffer 4, it is checked whether the input speech contains this ai word. Now in word buffer 4! ! It is assumed that the syllable string "ninshiki" of "L word r recognition" is stored.

The syllable candidate string generation section 5 creates a syllable candidate string from the syllable candidates in the syllable candidate storage section 2 in the order of the arrangement of syllables in the one-ton words stored in the word buffer 4, and combines the resulting syllable candidate string with the syllable candidate string. The corresponding syllable string is stored in the syllable candidate string storage section 6.

On the main stream side, a string of sounds ffi is created in order from the first syllable of single J1 (starting from 7).

First, since the first syllable of the word buffer 4 is "two", the syllable candidate string generation unit 5 is a syllable mata? +! In the i-storage unit 2, “
Extract the syllable candidates that have been classified into ``2'' and have been memorized.
Each of these is stored as a syllable candidate string of length 1 in the syllable mater supplementary sequencer α section 6 together with the syllable "two". As a result, the syllable candidate string storage section 6 stores two syllable candidate strings: ■2/2:4 (d) and ■2/10:13 (d). Here, the corresponding syllable string is in parentheses.

Next, the syllable candidate string generation unit 5 pays attention to the next syllable ``n'' and the next syllable ``shi'' in the M6D buffer 4, and determines whether they can be classified into ``n'' or ``shi'' in the syllable candidate storage unit 2. For each syllable candidate stored in the syllable candidate storage unit G
It is checked whether or not the last syllable candidate of any of the syllable candidates Yu U is read behind in the input speech. If there is a subsequent syllable candidate, that syllable candidate is connected to the end of the syllable candidate string to generate a new syllable candidate string and stored in the syllable candidate string storage section 6. The syllable candidate strings that are classified and stored as “n” or “shi” are n/2:4, n/13:16.
, senior/four-near.

Whether syllable candidate A reads behind another syllable candidate B can be determined by comparing the ending time of syllable candidate A and the starting end time of syllable candidate B. Here, it is determined that reading later is to be performed when the difference between these times is less than or equal to plus or minus 1. So in this case, the syllable candidate /13:
Connect 16 to the syllable candidate string ■ to make it correspond to the syllable string "nin", and connect the syllable candidate si/4 nia to the syllable candidate ■ to make it correspond to the syllable string 4 ninshi. The syllable candidate string that has been completely stored in the storage unit 6 is deleted.As a result, in the syllable candidate string storage unit 6, ■2/10:i3
Two syllable candidate sequences remain: -n/'13:16 ni n') ■2/2:4-shi/4 nia ni nn).

In the 6-syllable candidate storage unit 2, the syllable candidates that are classified and stored as "shi" or the next syllable "ki" are classified as "shi" and the next syllable "ki" are stored as "shi"/4-near. tree/
There are two at 16:19. For each of these, if we check whether the last syllable candidate of syllable candidate string ■ or ■ is read later in the input voice, Ki/16:19 is the sound 45 complement sequence ■
The last syllable candidate n/13: It is determined that it follows i6. As a result, the syllable candidate string ■ has a syllable candidate key of /16.
: A new syllable candidate string ■ created by concatenating syllables 19 and 19 is stored in the syllable string storage unit 6 in association with syllable number 11 "ninshiki".

Therefore, the contents of the syllable candidate string storage section 6 are as follows.

Here, since the last syllable in the word buffer 4 has been reached, the syllable candidate string generation unit 5 recognizes that the tun Bn1 recognition exists in the section from time 10 to time 19 in the input sound>'43. Don't output that.

(Above, one embodiment of the present invention has been described.) As long as the syllables are not consecutive, the syllables may be dropped multiple times in one nail word.

(Effects of the Invention) As explained above, according to the present invention, even if some non-consecutive syllables in a word to be detected are dropped at the stage of extracting syllable candidates from input speech, the presence of the word A word detection method that makes it possible to detect the position in input speech, and also allows for highly efficient word detection because the number of syllable candidate sequences generated during the detection process is extremely small. can be provided.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of input speech and extracted syllable candidates in the embodiment of FIG. 1, and FIG. 3 is a syllable in the embodiment of FIG. 1. It is a figure which shows an example of the content of a candidate storage part. 1... Syllable detection unit, 2... Syllable candidate storage unit, 3...
- Word storage section, 4... word buffer, 5... syllable string generation section, 6... syllable string storage section. Input audio Figure 1 Figure 2 Figure 3

Claims (1)

[Claims] By using a plurality of category candidates extracted from the input speech, which is a sequence of categories such as syllables, phonemes, and phoneme classes, and their position information, a category candidate sequence corresponding to the word category sequence is generated. In a word detection method that detects words and their positions in speech, multiple category candidates obtained from input speech are classified and stored by their category names, and each category is sorted according to the order of the categories in the word. Select the category candidate corresponding to the category from among the stored category candidates classified with the same name as the category, and select the first and last categories of the three consecutive categories in the word in the input speech. When corresponding to each category candidate in a sequence of two consecutive category candidates, the sequence of three categories is made to correspond to the sequence of two category candidates to generate a category candidate sequence. A word detection method that uses