JPH0632014B2 - Word detection method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a word detection device, and more particularly to a word detection device that detects a word included in an input voice in a voice recognition device, a voice input device, or the like.

(Prior Art) In a voice recognition device, a voice input device, etc., as one of the methods for detecting words in the input voice, the input voice is regarded as a sequence of categories of syllables, phonemes, and phoneme classes and detected from the input voice. There is a method in which if a category string created using a category corresponds to a category string of a word stored in a word dictionary, that word is used as a detection result.

In general, it is difficult to detect only the correct category from the input speech without error because the utterance time length of the above-mentioned category is short and similar categories exist. Therefore, conventionally, a method of detecting a plurality of category candidates for each category section in the input speech and then finding a category candidate string corresponding to a word by referring to a word dictionary has been used. However, also in this case, the existence of a certain category section may not be detected or a correct category candidate may not be detected in the category section due to the cause of the voicing bluntness or adjacent categories such as articulation of syllables.

Therefore, Japanese Patent Application No. 61-190258, 190259, 1
As in the method described in “Word detection method” of 90260, 190261, after detecting the category candidate in the input speech, the category candidate is selected according to the category arrangement of the words in the word dictionary, and a certain category in the word is selected. If is not detected, the category candidates corresponding to the categories before and after that are used as clues to find the category candidate string corresponding to the word category string. Further, regarding the word for which the corresponding category candidate string is found, the score of the category candidate string is calculated and used as the word score.

On the other hand, when selecting a word with a good score from a large number of words, if the corresponding category candidate string and score are individually obtained for each word by the above method, the number of words in the word dictionary is particularly large. In that case, it requires a large amount of calculation and is not realistic.

Therefore, by preferentially obtaining the category candidate sequence having a good score, the word having a good score is obtained without obtaining the corresponding category candidate sequence for all the words. That is, the category string of a word is traced from its start end to obtain a category candidate sequence corresponding to a part of the way up to that point, and then the scores of the category candidate sequences up to that point are obtained. Then, at each time point, the category candidate sequence having the best score is selected, and the corresponding category sequence is further traced. When the end of the word which is repeating this process is reached, that word is output as a result. As a result, the word corresponding to the category candidate sequence having a good score is preferentially obtained.

Here, as the score of the category candidate string, the average value of the scores of the category candidates forming the category candidate string is always given.

(Problems to be Solved by the Invention) As described above, when the score of the category candidate sequence corresponding to the beginning to the middle of the category sequence is obtained from the average value of the scores of the category candidates forming the category candidate sequence, When the score of the candidate corresponding to the category close to the beginning of the category sequence of the correct word is much worse than the other candidates, there is a drawback that the word is often not detected easily.

FIG. 3 is an explanatory diagram showing an example of voice candidates extracted from the input voice. For example, using syllables as categories,
It is assumed that when the voice "Zaysei" is input, a syllable candidate as shown in FIG. 3 is obtained for each syllable. Here, the number written in each syllable candidate is the score of the syllable candidate, and the smaller the value, the better, that is, the more reliable. At this time, the syllable candidate sequence corresponding to each of the words “seigen (restriction)”, “seigen (finance source)”, “zeisei (tax system)”, and “seisei (financial)” is calculated using the syllable candidates shown in FIG. When created, the score is: Seigen: (3 + 7 + 8 + 4) /4=5.5 Seigen: (8 + 7 + 8 + 4) /4=6.75 Zeisei: (12 + 7 + 1 + 1) /4=5.25 Seisei: (8 + 7 + 1 + 1) /4=4.25 And the correct word "Zaysei" has the lowest score and therefore the highest score. However, the length from the beginning to the middle of the syllable string of each of these words is n.
The score of the syllable candidate sequence corresponding to the (n = 1,2,3,4) partial syllable sequence is calculated as the average value of the scores of the respective syllable candidates as follows.

Center: 3/1 = 3 Say: (3 + 7) / 2 = 5 Sayage: (3 + 7 + 8) / 3 = 6 Seigen: (3 + 7 + 8 + 4) /4=5.5 The: 8/1 = 8 Zay: (8 + 7) / 2 = 7.5 Zaige: (8 + 7 + 8) /3=7.66 Zeigen: (8 + 7 + 8 + 4) /4=6.75 Ze: 12/1 = 12 Zei: (12 + 7) /2=9.5 Zeise: (12 + 7 + 1) / 3 = 6.66 Zeise: (12 + 7 + 1 + 1) /4=5.25 The: 8/1 = 8 Zei: (8 + 7) /2=7.5 Zeise: (8 + 7 + 1) /3=5.33 Zeisei: (8 + 7 + 1 + 1) /4=4.25 Therefore, when the syllable string is traced while selecting the syllable candidate string having the best score as described above, the syllable string of the above four words is traced in the following order. Where a colon (:)
The number on the right is the score of the syllable candidate sequence corresponding to that syllable sequence.

First, a syllable candidate string corresponding to the leading syllable of each of these four words is obtained.

S: 3 The: 8 The: 8 Z: 12 Next, the process of selecting the syllable candidate sequence having the best (smallest) score among these and tracing it to the destination is repeated. Then, the syllable candidate sequence is created in the following order.

Say: 5 Sayge: 6 Saygen: 5.5 Zay: 7.5 Zayge: 7.66 Zaygen: 6.75 Zay: 7.5 Zaysei: 4.25 Thus, the syllables of the words “Saigen” and “Zaigen” By obtaining the syllable sequence of the word "Zaisei" for the first time after finishing the sequence, the wrong words "limit" and "finance source" are detected before the correct word "finance". This is because the score of the syllable candidate corresponding to the first "Za" in "Zaisei" is bad, so "Za" of "Saigen" is not selected until there is no syllable candidate string with a score better than 8. To do.

The object of the present invention is to eliminate the above-mentioned drawbacks, and thus, even when the score of the candidate corresponding to the category close to the start of the category sequence of the correct word is significantly worse than the other candidates, the correct word is further forwarded. Another object of the present invention is to provide a word detection method that enables detection with a small amount of processing.

(Means for Solving Problems) The word detection method of the present invention is a method of detecting a plurality of category candidates from an input speech that is a sequence of categories such as syllables, phonemes, and phoneme classes, and reliability in detection evaluation of these category candidates. Detecting words in the input speech by extracting the position information of the score as a measure and finding the category candidate string corresponding to the word category string while tracing the word category string stored in the word dictionary In the method, when the score of a category candidate string consisting of n category candidates is n or more, which is a predetermined number N, the average value of the scores of the n category candidates is used to calculate n.
When it is less than, it is calculated by using the average value of the scores of (n + m) category candidates obtained by adding n virtual category candidates as a preset function depending on n and N to n category candidates. And a means for tracing the category sequence of words in the word dictionary so as to always obtain the category candidate sequence having the best score.

(Operation) In the above example, although the syllable candidate sequence corresponding to the syllable sequence “Zaysei” has a score of 4.25, which is better than the scores of other words, it corresponds to the leading syllable “The”. The score of the syllable candidate who does is bad as 8. On the other hand, the syllable candidate sequence corresponding to the syllable sequence “Saigen” has a score of 5.5, which is worse than that of the word “Zaisei”, but the score of the syllable candidate corresponding to the leading syllable “SE” is 3 and the word “Zaisei”. It is better than the score of the first syllable candidate.

As described above, the reliability of a word is correctly evaluated as a score calculated from the category candidate string corresponding to the category string of the entire word, but in the conventional technique, only the category candidate string corresponding to the category string of a part of the word is evaluated. Since the score calculated from is treated as the score of the word, the score of the word becomes worse when a part of it accidentally contains a category candidate having a bad score.

On the other hand, when the category candidate string corresponding to only the category string of a part of the word is not determined, the score of the category candidate string corresponding to the category string of the entire word cannot be used.

Therefore, in the method of the present invention, among the categories in the word category string, the category candidate having a certain average score is hypothesized for the category for which the corresponding category candidate has not been determined yet. That is, when the score of a category candidate string is obtained, if the length n is shorter than a predetermined length N, it is determined that the category candidate string corresponds to a category string that is a part of a word. , N and N
The number of category candidates is hypothesized as a function that depends on and is set in advance. As a result, even when a part of the corresponding category candidate string includes a category candidate having a bad score, the score of the word is not so bad because the category candidate string is averaged by the score of the virtual category candidate string. Therefore, the category candidate sequence corresponding to the word can be quickly obtained.

(Example) Next, this invention is demonstrated in detail with reference to drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. In the embodiment shown in FIG. 1, Japanese voices are input, and syllables are used as categories.

The syllable candidate extraction unit 101 detects a syllable candidate in the input voice and stores the candidate in the syllable candidate storage unit 102 together with the position of the score in the input voice.

FIG. 2 is a block diagram showing an example of the syllable candidate extraction unit 101. In FIG. 2, the input voice is the voice puffer 20.
It is temporarily stored in 1. First, the vowel candidate detection unit 202 detects a vowel candidate in the voice stored in the voice buffer 201 and stores it in the vowel candidate storage unit 203. The vowel candidate is detected by comparing the voice standard pattern of each vowel previously stored in the vowel pattern storage unit 204 with each section of the input voice. The vowel voice signal is relatively stationary and therefore easy to detect. Each vowel holds at least the vowel name and position information in the input voice.

After the detection of the vowel candidate is completed, the consonant candidate detection unit 205 detects the consonant candidate as described below. In Japanese, a syllable is a consonant (C) -vowel (V) pair. Therefore, in the input speech, a section within a certain time length (hereinafter referred to as a VCV section) of a section sandwiched between two vowels and a section within a certain time length from the beginning of the input speech (hereinafter referred to as a VCV section). It can be said that there is one consonant in each of up to the CV section). The consonant candidate detection unit 204 stores VCs stored in advance in the consonant pattern storage unit 206 for all VCV sections and CV sections created from the vowel candidates stored in the vowel candidate detection unit 203.
The V and CV standard voice patterns are collated, and the names of a plurality of voice patterns having a high degree of similarity are used as consonant candidates. The vowel candidate and the consonant candidate determined above are combined into a syllable candidate, and the syllable candidate storage unit 10 together with the position in the input voice.
Store in 2.

As an example, it is assumed that the voice "Zaisei (Finance)" is input. In this case, syllable candidates such as those shown in FIG. 3 are extracted as the syllable recognition result. In FIG. 3, a plurality of syllable candidates are extracted in each syllable section, and the number described in each syllable candidate is the score of the candidate.

The word storage unit 103 stores syllable strings of words to be detected. Now, in the word storage unit 103, "seigen (restriction)", "seigen (finance resource)", "zeisei (tax system)"
It is assumed that the four words “Zaysei (Finance)” are stored.

The syllable candidate string generation unit 104 first obtains the syllable candidate corresponding to the beginning of each word in the word storage unit 103 as the syllable candidate storage unit 10.
The syllable candidates stored in No. 2 are selected, and each is made a syllable candidate string of length 1. Next, the score calculation unit 105 calculates the score of each syllable candidate string. In this embodiment, N = 4, m as a function preset depending on n and N uses m = N−n, and the score of the virtual syllable candidate is 1.

For example, the score of a syllable candidate sequence consisting of only the syllable candidate sequence “3” corresponding to the first syllable “S” of the word “Saigen” has a total of 4 virtual syllable candidates including this syllable candidate and 3 scores 1. Since it is the average value of the scores of the individual syllable candidates, (3
+ 1 + 1 + 1) /4=1.5.

These syllable candidate strings are stored in the syllable candidate string storage unit 106 together with the words corresponding to the scores. As a result,
The syllable candidate string storage unit 106 stores four syllable candidate strings of C: 1.5 “Saigen” The: 2.75 “Zaigen” Z: 3.75 “Zeisei” The: 2.75 “Zaisei” Has been done. Here, the inside of "" is a corresponding word.

Next, the syllable candidate string selection unit 107 determines the syllable candidate string storage unit 1
Among the syllable candidate sequences in 06, the syllable candidate sequence having the best score, that is, the smallest value is taken out, and the syllable candidate sequence and the word are sent to the syllable candidate sequence generation unit 104. The syllable string candidate string generation unit 104 further extends the received syllable string candidate according to the syllable string of the word, calculates the score again by the score calculation unit 105, and then stores the score in the syllable candidate string storage unit 106.

In the present case, SE: 1.5 “SEIGEN” is taken out, and SEI: 3 “SEIGEN” is newly stored in the syllable string storage unit 106. As a result, the syllable candidate string storage unit 106 stores the: 2.75 “Zaygen” Ze: 3.75 “Zaysei”: 2.75 “Zaysei” Sei: 3 “Saigen”. Therefore, next, the syllable candidate sequence having the best score among these, the: 2.75 "Zeigen", is extracted.

By repeating the above process, the following syllable candidate sequence is created.

ZAI: 4.25 “Zaygen” ZAI: 4.25 “Zaysei” SAIGE: 4.75 “SAIGE” ZAIGE: 6 “ZAIGEN” ZISE: 4.25 “ZAIsei” ZAISEI: 4.25 “ZAISEI” where syllables Since the candidate string has reached the end of the word “Zaysei”, the syllable candidate string selection unit 107 outputs this word as a detection result.

Thus, the correct word "Zaysei" is first detected. Moreover, in the conventional method, "Zaysei" is detected when 13 syllable candidates are created as described above, whereas in the method of the present invention, 11 syllable candidate strings are detected. .

The embodiment of the present invention has been described above. In this embodiment, for simplification of description, the case where no syllable recognition error occurs in the stage of syllable recognition, that is, the case where at least correct syllable candidates are extracted for all input syllables is described. Even when a recognition error occurs, the aforementioned Japanese Patent Application No. 61-190258, 190259, 190
By using the method described in the "word detection method" of 260, 190261, a correct word can be detected efficiently as in the above embodiment.

When many words to be detected exist, they are expressed in a tree structure format, that is, syllables are used as nodes, and the node string from the root node to the leaf node represents the syllable string of each word. However, in this case also, regarding each syllable string, when the method of the present invention is applied, the same result as the above-mentioned embodiment is obtained.

(Effects of the Invention) As described above, according to the present invention, even when the score of a candidate corresponding to a category close to the start end of the category sequence of a correct word or when the score is significantly worse than other candidates, the correct word is excluded. There is an effect that it is possible to realize a word detection method that can detect words earlier than, and generate a small number of category candidate strings, and can perform efficient word detection.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a block diagram showing an example of a syllable candidate extraction unit in the embodiment of FIG. 1, and FIG. 3 is a syllable candidate extracted from an input voice. It is explanatory drawing which shows an example. 101 syllable candidate extraction unit, 102 syllable candidate storage unit, 103 word storage unit, 104 syllable candidate string generation unit, 105 score calculation unit, 106 syllable candidate string storage unit, 107 ... Syllable candidate string selection unit, 201 ... Voice buffer, 202 ... Vowel candidate detection unit, 203 ... Vowel candidate storage unit, 204 ... Vowel pattern storage unit, 205 ... Consonant candidate detection unit, 206 ... Consonant pattern Memory.

Claims (1)

[Claims] 1. A plurality of category candidates and a score and position information as a measure of reliability in detection and evaluation of these category candidates are extracted from an input speech that is a sequence of categories such as syllables, phonemes, and phoneme classes, and words are extracted. In a word detection method for detecting a word in an input voice by finding a category candidate string corresponding to a word category string while tracing a category string of words stored in a dictionary, a category candidate string consisting of n category candidates Is calculated by using the average value of the scores of n category candidates when n is a predetermined number N or more, and when n is less than N, a function that is preset depending on n and N for the n category candidates Is calculated by using the average value of the scores of (n + m) category candidates including m virtual category candidates as A word detecting method comprising means for tracing a category string of words in the word dictionary so as to always obtain a category candidate string having the best score.