JPS60149096A - Recognition of word voice - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a word speech recognition method for recognizing words by comparing input speech with a word dictionary in which phonemes are expressed.

(Constitution of Conventional Example and Problems thereof) FIG. 1 is a functional block diagram of an apparatus for executing an example of a conventional word speech recognition method and an embodiment of the word speech recognition method of the present invention. A conventional example will be explained with reference to FIGS. 1 to 3. In Figure 1, 1 is a parameter extraction unit that creates a time series of parameters from input speech, 2 is a probability density calculation unit that calculates the probability density of a phoneme by collating a phoneme standard pattern, and 3 is a probability density calculation unit that calculates the probability density of each phoneme. segmentation, likelihood calculation,
il' is a word recognition unit that performs four-word similarity calculations and the like. 1
4 is the average value (/1
11), and a phoneme standard/zotan part that stores phoneme standard patterns expressed in the form of covariance matrices (Σ,) between various parameters; 5 is a word in which all words to be recognized are expressed as symbol strings in phoneme units; This is a word dictionary section in which dictionaries are stored. In the word dictionary, for example, the words "su77e 口J, r ASAHIKAWA", "Akita", "shima", "shisa", etc. are respectively rsAQPOROJ, rASAHIKAWA J,
It is written as rAKITAJ, rsIMAJ, "5ISA", etc.

Next, the operation of the above conventional example will be explained. The input phoneme is analyzed by the parameter extraction unit 1 for each frame of 10 m5, and parameters are extracted to create a parameter time series.

The probability density calculation unit 2 compares the grammeter obtained for each frame with a phoneme standard pattern, and calculates the probability density of the phoneme generated from the value of the grammeter. Next, in the word recognition unit 3, using the above parameters and the obtained probability density value, phoneme segmentation is performed for each dictionary item according to the dictionary phoneme sequence that constitutes the dictionary item, and the following 0 formula is used. Then, the type of phoneme and the likelihood of the segmented interval corresponding to the phoneme are calculated, and the similarity is calculated as the average of the likelihoods of each phoneme in the dictionary entry. Here, let the phoneme be X, and the frame numbers at the start and end of the segmented section corresponding to X are N5°Ne! :L, and when the value of each parameter in the n-th frame is Cn, the likelihood LX of phoneme X is defined by the following formula.

φ1(Cn) represents the probability density of a certain phoneme 1, and is defined as in equation (2).

・・・・■C1 J] fixed parameter (vector) in one frame sti: Mean value (vector) of the parameter of a certain phoneme l Σ,: Covariance matrix The range of summation l differs depending on the phoneme X. For example, when X is the phoneme A (g), the range of l is 5 vowels, A
, , E, I, O, U.

The word similarity LM obtained above was calculated for each dictionary item according to formula (2), and the dictionary item with the maximum LM was selected as a recognized word.

LM: Similarity of the Mth word in the dictionary ti: Likelihood of phoneme i in the dictionary phoneme sequence NP: Number of labial phonemes In the above conventional example, one phoneme in the dictionary entry is calculated using the probability density value of the phoneme. Segmentation and likelihood calculations are performed for each. FIG. 2 shows the temporal change in the probability density of each phoneme when /SiMA/ (island) is uttered. In this case, the segmentation and likelihood calculation are performed using the probability density value φ8 of each phoneme /S/, /i/, /M/, /A/.
．． Segmentation of /S/ at the beginning of the word is performed according to the time changes of φl, φM, and φA. The segmentation of /S/ at the beginning of the word is performed in the frame where φ8 is low and φi is high, and a is the rear end of /S/, and the segmented section (SF-a) Likelihood calculation is performed using φ8 for. Similarly, for the second phoneme /i/ following the /S/ at the beginning of the word, frame b where φi is low and φM is high is set as the rear end of 4v, and φi is set for the segmented interval (a-b). Likelihood calculations were performed using .

Figure 3 shows the temporal change in the probability density of each phoneme when /5iSA/ (suggestion) is uttered. Segmentation and likelihood calculation are performed for each phoneme /Sl/, A /, /s/
.

The probability density value of /A/, φ8. φ1. φ8. This is done according to the time change of φ^, but when segmenting the word-initial /S/, φ1 becomes very small because the following /1/ is devoiced, and φ8 is the original segment of the word-initial /S/. (FS=c), and the /S at the beginning of the word
Since the original interval (c-d) of /1/ that follows / is also exceeded, the trailing end e of /S/ that follows /i/ is changed to /S at the beginning of the word.
/ was output as the trailing end, causing a segmentation error.

Therefore, the phonemes /i /, /s/, following the /S/ at the beginning of the word,
The segmentation of the /A/ phoneme is also incorrect, and as a result, words containing devoiced vowels are easily misrecognized as a result of the lower likelihood.

(Objective of the Invention) The present invention aims to eliminate the drawbacks of the above-mentioned conventional example, and aims to improve the accuracy of segmentation and likelihood calculation, thereby improving the word recognition rate.

(Structure of the Invention) The present invention includes a word dictionary in which words to be recognized are expressed as symbol strings in units of phonemes, and a mark W-4' tongue for each phoneme expressed in the distribution form of acoustic parameters of each phoneme. When recognizing words in input speech, the input speech is checked against each dictionary entry in the word dictionary, and the phoneme standard is used for each phoneme according to the dictionary phoneme sequence that makes up each dictionary entry. The input speech is segmented by calculating the probability density generated from the above, and words are recognized by calculating the similarity between each dictionary item and the input speech using the above probability density value for the segmented speech section. In a word speech recognition method, when performing segmentation and likelihood calculation of a devoiced vowel sandwiched between unvoiced consonants, the probability density value of each phoneme is used to identify unvoiced consonants and unvoiced vowels, including devoiced m-sounds. This method is characterized by segmenting three consecutive phonemes of unvoiced consonants and performing likelihood calculations, which has the effect of improving the accuracy of segmentation and likelihood calculations.

(Description of Embodiment) An embodiment of the present invention will be described below with reference to FIG. 1. In the same figure, the parameter extraction section 1, the probability density calculation section 2, and the phoneme standard/zoom section 4 are the same as those in the conventional example described above. This is part of the segmentation and likelihood calculation in Section 3.

The word dictionary stored in the word dictionary section 5 has 44 words to be recognized (expressed as phoneme symbol strings), but the difference from the conventional example is that 1 vowel that is easy to be devoiced, such as "AsAH
■KAWA J% I” AK (1) TAJ, r
S a) MA J , [S oS A J , etc.) 11 marked with an O must be given a symbol to indicate this in advance.

The method of this embodiment first obtains a parameter for each frame from the input speech using the ξ parameter extraction unit 1, and then uses the value of the parameter in the probability density calculation unit 2 to calculate the probability density obtained from each phoneme standard button. Calculate. The process up to this point is the same as the conventional example. Next, the word recognition unit 3 performs segmentation of the phoneme X according to the dictionary phoneme series that constitutes the dictionary item for each dictionary item in the word dictionary unit 5. Calculate the likelihood tx. When there is a devoiced vowel V between voiceless consonants gC,,C2 in the dictionary phoneme sequence, the probability density value of the devoiced vowel does not indicate the nature of the vowel.
Indicates the nature of voiceless consonants. Therefore, in the above segmentation, voiceless consonants, voiceless vowels, voiceless consonants (C-1V
Segment the three phonemes at once using the type of each phoneme in the sequence C2), the phoneme, and the corresponding phoneme probability density value, and calculate the likelihood 'c, Calculate vc2.

In Figure 3, φ1 of the probability density of /V between /sis/
Instead, the probability density value φS of 'S/ at the beginning of the word is dominant from the beginning of the word to the end of the third S, e.

Therefore, the probability density of the third phoneme /S/ of three consecutive phonemes including a devoiced vowel and the vowel that follows it, the sound /A/, is φ8.
Segmentation is performed using φA, and likelihood is calculated using φ8 for the segmented interval. By doing this, /SIS/, a series of three phonemes consisting of a voiceless consonant, a voiceless vowel, and a voiceless consonant, becomes an interval (FS
-e), and good segmentation is possible, which improves the accuracy of likelihood calculation.

In this example, the devoiced vowel is not treated as one phoneme, but the phoneme sequence of the unvoiced consonant, devoiced vowel, and unvoiced consonant including the devoiced vowel is collectively calculated to perform segmentation likelihood, so the devoiced U sound is This has the advantage of improving the recognition rate of words included.

(Effects of the Invention) The present invention uses the probability density value of each phoneme when performing segmentation and likelihood calculation of a devoiced m-meter that is suitable for unvoiced consonants. conjugated vowel,
Since three consecutive phonemes of unvoiced consonants are segmented and likelihood calculations are performed collectively, this method has the advantage of performing segmentation and likelihood calculations with higher precision than conventional methods.

[Brief explanation of drawings]

FIG. 1 is a functional block diagram of an example of a conventional word speech recognition method and an embodiment of the word speech recognition method of the present invention, and FIG. 2 is a case in which /SiMA/, (island) is uttered. FIG. 3 is a diagram showing changes in the probability density of each phoneme over time when /5iSA/ (suggestion) is uttered. 1... Parameter extraction unit, 2... Probability density calculation unit, 3
・・・Word recognition part, 4... Phoneme standard slam part, 5...
Word dictionary department. Figure 1 Input scream

Claims (1)

[Claims] jl'+, which represents the word to be recognized as a symbol string in phoneme units
It is equipped with a word dictionary and a standard pattern for each phoneme expressed as a distribution form of the acoustic parameters of each phoneme. According to the dictionary phoneme series that constitutes the top of each dictionary, the probability density generated from the phoneme is calculated for each phoneme using the phoneme standard Gutan, the input speech is segmented, and the segmented speech interval is calculated. On the other hand, when recognizing words by calculating the similarity between each dictionary item and input speech using the above probability density value, for a devoiced vowel narrowed to a voiceless consonant, a devoiced vowel is including, voiceless consonants,
A method for recognizing idiographic speech, characterized in that continuous consonants are segmented using probability density values of phoneme sequences of devoiced vowels and unvoiced consonants, and likelihood calculations are performed.