JPH0656556B2 - Word detection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention is used in a voice recognition device, a voice input device, etc., and is a word for detecting a word included in an input voice and a position of the word in the voice. Regarding detection method.

(Prior Art) A method for detecting a word and its position in an input voice in a voice recognition device, a voice input device, etc., includes a method for inputting each category from the input voice that is a sequence of categories such as syllable, phoneme, and phoneme class Position information in the voice is extracted, and if the category string created from the extracted categories corresponds to the category string of a word, that word and the position of the category string in the input voice are output as the detection result. There is a way.

In general, the above-mentioned categories have a short time length, and there are similar categories. Therefore, it is difficult to extract the categories in the input speech completely without error. Therefore, conventionally, a plurality of category candidates are extracted for each category section in the input voice, and partial category candidate strings are generated using the category candidates in order from the end of the input voice. Found a category candidate sequence corresponding to the word by repeating the process of matching the word with the category sequence. Details of this method are described in, for example, Japanese Patent Application No. 58-214544, Pattern Recognition Device, and will not be described here.

Also, at the stage of extracting categories in the input voice, a certain category has disappeared or its presence could not be detected due to factors such as the occurrence of occurrence and articulatory coupling between adjacent categories (for example, syllables). The categories before and after that may be extracted as being adjacent to each other. This phenomenon is hereinafter referred to as category dropout.

In order to deal with this, in the past, it was investigated in advance what sort of categories would result in category dropouts, and from the results, the category dropouts for relatively frequent category dropouts have occurred. Prepare a category string correction rule to convert a category that is missing from the category string into a restored category string. By applying this rule to the category candidate sequence at the time of word detection,
For a relatively frequently dropped category, the dropped category can be restored. The details of this method are described in, for example, Reference 2 “Namiki, Hamada, Nakatsu,“ Japanese Input Method Using Speech Recognition ”, SIJ Vol. J67-D No. 4, 1984.
No. 4 ”is omitted here.

(Problems to be Solved by the Invention) In the above-mentioned conventional method, since the category candidate string is generated using the category candidates extracted from the input speech, the category candidate string is collated with the word category string. A large number of category candidate sequences, which are wasted, are generated, which requires a large amount of calculation.

Further, even if a word segment to be detected occupies only a part of the input voice, conventionally, all the category candidates including the segment in which the word does not exist are matched equally with the category in the word from the end of the input voice. Must be done,
It takes a lot of calculation time and a long time to detect a word.

Further, the above-mentioned category string correction rule is applied not only to the category candidate string in which the category dropout has occurred but also to the category candidate string in which the category dropout has not occurred. In many cases, a plurality of correction rules are individually applied to one category candidate sequence. Therefore, many category candidate sequences are generated from one category candidate sequence,
Many category candidate sequences must be examined until a category candidate sequence corresponding to the word to be found is found.
Moreover, most of the category candidate sequences do not match the category sequence of the word to be detected, and are eventually wasted.

Further, the categories that can be restored by the correction rule are limited to the categories that are relatively frequently dropped, and the categories that are relatively rarely restored cannot be restored. The number of correction rules must be increased in order to increase the number of dropouts that can be restored, and as a result, more and more category candidate sequences are generated.

For example, when trying to extract a syllable candidate from a voice generated as "Onsei Ninshitowa (What is voice recognition)", the duration of the syllable "Si" is short and the syllable "Si" and its Since the syllable boundaries of the preceding and following syllables "n" and "ki" were close to each other, the existence of the syllable "si" could not be detected, and the syllable candidates of the preceding and following syllables "n" and "ki" were detected. It is assumed that the data are extracted at the adjacent positions. As a result, even if correct syllable candidates are obtained for all syllables other than the syllable "si", the generated syllable candidate sequence becomes "onseininkittowa", and the dropped syllable "si" is corrected by the correction rule. Must be restored at. However, such syllable omissions are a relatively rare type, and it seems unlikely that there will be rules to correct these omissions. Also, even if this correction rule is prepared,
In addition to this, for example, "nin" → "nine", "a" →
In many cases, a correction rule called "Good" is prepared,
By applying these, useless syllable candidate sequences such as “onseiinnintowa”, “onseiinnintowa”, and “onseiinnintowa” are generated.

An object of the present invention is not to generate a wasteful category candidate sequence, and when a section of a word to be detected is a very small part of the entire input speech, or when some of the words to be detected in the input speech are detected. An object of the present invention is to provide a word detection method capable of efficiently detecting a word and its position from an input voice even when the category of is dropped.

(Means for Solving Problems) Means provided by the present invention in order to solve the above problems and achieve the above object are extracted from an input speech which is a sequence of categories such as syllables, phonemes, and phoneme classes. A word detection method for detecting a word in an input voice and its appearance position by generating a category candidate string corresponding to a word category string using a plurality of category candidates and their position information, A plurality of category candidates obtained from the input voice and their position information are classified and stored by the category name, and the category candidates and their position information corresponding to each category are stored according to the order of the categories in the word. While selecting from the category candidates that are categorized into the same name as the category and stored, the first and last categories in the sequence of three consecutive categories in a word are the input sounds. When corresponding to each category candidate in the sequence of two consecutive category candidates in the voice, the sequence of three categories and the sequence of two category candidates are associated with each other to generate a category candidate sequence. It is characterized by performing.

(Operation) In the method of the present invention, among the category candidates extracted from the input speech, only the category candidates having the same name as the category included in the word to be detected are used, and the arrangement of the categories in the word is followed. A category candidate sequence to be generated is generated. As a result, only the category candidate string corresponding to the word category string or the substring thereof is generated, and it is possible to avoid generating an unnecessary category string.

In addition, in order to generate a category candidate string from category candidates corresponding to the category in a word among the category candidates in the input speech, when the section of the word to be detected is a very small part of the entire input speech. It is possible to detect the word quickly regardless of whether or not the section exists in the input voice.

If a category in a word in the input speech is dropped,
The category candidates for the three adjacent categories before and after the category are adjacent to each other in the input voice.
That is, the category string in a word is set to C _i-1 C _i C _{i + 1} ,
When the category C _i is dropped, each category candidate is extracted as C _{i + 1} follows C _i-1 .

Therefore, when the category candidate string corresponding to the word is generated while tracing the arrangement of the categories in the word to be detected, the category candidate corresponding to C _{i + 1} is the category candidate corresponding to C _i-1. If it follows in the input speech, the sequence of the two category candidates is the category string C _{i-1 in the} word.
Corresponds to C _i C _{i + 1} . As a result, even if the category C _i is missing, it is possible to take a correct correspondence between the category candidate sequence and the category sequence. Further, since only the category candidate string corresponding to the word category string is generated, useless generation of the category candidate string can be avoided.

(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 an embodiment of the present invention.

In the present embodiment, syllables are used as categories assuming that Japanese voice is input.

The syllable extraction unit 1 detects syllable candidates in the input voice and stores the candidates in the syllable candidate storage unit 2. As an example, it is assumed that the voice “ON-SEI NIN-SHITOH” (what is voice recognition) is input. In this case, as a result of syllable recognition, for example, syllable candidates as shown in FIG. 2 are extracted. In FIG. 2, a line of an arrow is a section of each syllable candidate, and a plurality of syllable candidates are extracted in each section. These syllable candidates are classified by syllable name and stored in the syllable candidate storage unit 2. As a result, the contents of the syllable candidate storage unit 2 are as shown in FIG. In this figure, each syllable candidate is represented by “syllable name /
It is expressed in the format of "start time: end time".

The word storage unit 3 stores syllable strings of words to be detected. After extracting one word from the word buffer 4, it is examined whether or not this word is included in the input speech. Now, it is assumed that the word buffer 4 stores the syllable string “Ninshiki” of the word “recognition”.

The syllable candidate string generation unit 5 creates a syllable candidate string from the syllable candidates in the syllable candidate storage unit 2 in the order of arrangement of syllables in the words stored in the word buffer 4, and corresponds to the resulting syllable candidate string. The syllable sequence and the syllable candidate sequence storage unit 6 are stored. In this embodiment, a syllable string is created in order from the beginning syllable of a word.

First, since the first syllable of the word buffer 4 is “d”, the syllable candidate string generation unit 5 stores “d” in the syllable candidate storage unit 2.
The syllable candidates that have been classified and stored are extracted and stored as syllable candidate strings of length 1 in the syllable candidate string storage unit 6 together with the syllable "d". As a result, the syllable candidate string storage unit 6 stores two syllable candidate strings of N / 2: 4 (D) N / 10: 13 (D). Here, the brackets are the corresponding syllable strings.

Next, the syllable candidate string generation unit 5 pays attention to the next syllable “n” and the next syllable “si” in the word buffer 4 and classifies them into “n” or “si” in the syllable candidate storage unit 2. For each of the stored syllable candidates, it is checked whether or not the last syllable candidate of any syllable candidate sequence in the syllable candidate storage unit 6 is followed in the input speech. If there is a succeeding syllable candidate, the 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 unit 6. The syllable candidate strings that are classified and stored as “n” or “si” are n / 2: 4, n13: 16, and si / 4:
3 of 7. Whether or not the syllable candidate A follows the other syllable candidate B can be determined by comparing the end time of the syllable candidate A and the start time of the syllable candidate B. Here, when the difference between those times is plus or minus 1 or less, it is determined to follow. Therefore, in the present case, syllable candidate / 13: 16 is connected to the syllable candidate sequence to correspond to the syllable sequence “nin”, and syllable candidate sequence / 4: 7 is connected to the syllable candidate to produce the syllable sequence “ninshi”. Correspond to. Further, the syllable candidate string stored in the syllable candidate string storage unit 6 up to that point is deleted. As a result, in the syllable candidate string storage unit 6, two syllable candidate strings of N / 10: 13-N / 13: 16 (Nin) N / 2: 4-Si / 4: 7 (Ninshi) are stored. Remain.

Then, the process proceeds to the syllable "si". In the syllable candidate storage unit 2, the syllable candidates that are classified and stored as “si” or the next syllable “ki” are si / 4: 7 and ki / 1.
It is two of 6:19. For each of these, it was determined whether or not the syllable candidate sequence or the last syllable candidate of the syllable candidate sequence was followed in the input voice, and it was determined that Ki / 16: 19 follows the last syllable candidate / 13: 16 of the syllable candidate sequence. It As a result, a new syllable candidate string formed by connecting the syllable candidate key 16:19 to the syllable candidate string is stored in the syllable string storage unit 6 in association with the syllable string “Ninshiki”. Therefore, the contents of the syllable candidate string storage unit 6 are 2/10: 13- / 13: 16-key / 16: 19 (ninshinki).

Here, since the last syllable in the word Bazza 4 has been reached, the syllable candidate string generation unit 5 outputs that the word “recognition” exists in the section from time 10 to time 19 in the input speech. To do.

The embodiment of the present invention has been described above. It should be noted that, as long as the syllables are omitted, a plurality of syllables may occur in one word unless they are consecutive.

(Effects of the Invention) As described above, according to the present invention, even if some non-consecutive syllables in a word to be detected are dropped at the stage of syllable candidate extraction from an input voice, the existence of that word exists. And a position in the input voice can be detected, and the number of syllable candidate sequences generated during the detection process is extremely small, which enables efficient word detection. Can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of an input voice and extracted syllable candidates in the embodiment of FIG. 1, and FIG. 3 is a syllable in the embodiment of FIG. 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 unit, 4 ... word buffer, 5 ... syllable string generation unit, 6
...... Syllable string storage unit.

Claims (1)

[Claims] 1. A category candidate sequence corresponding to a word category sequence is generated using a plurality of category candidates extracted from an input speech, which is a sequence of categories such as syllables, phonemes, and phoneme classes, and their position information. By doing so, in a word detection method for detecting a word in an input voice and its appearance position, each of a plurality of category candidates and their position information obtained from the input voice is classified and stored by the category name, and the word is stored. The category candidates corresponding to each category and their position information are selected from the category candidates stored under the same name as the category according to the order of arrangement of the categories, and three consecutive categories in the word are selected. When the first and last categories in the sequence of correspond to the respective category candidates in the sequence of two consecutive category candidates in the input speech, A sequence of three categories and the alignment of two categories candidates in correspondence, features and to word detection scheme to perform the generation of the category candidate string.