JPH045391B2 - - Google Patents

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 example of a conventional word speech recognition method and an apparatus for carrying out an embodiment of the word speech recognition method of the present invention. Conventional example first
This will be explained with reference to FIGS. 2 and 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 comparing standard patterns of phonemes, and 3 is a segmentation unit for each phoneme. This is a word recognition unit that performs similarity calculations, likelihood calculations, and word similarity calculations. 4 is a phoneme standard pattern section that stores a phoneme standard pattern representing the distribution of various parameters for each phoneme in the form of an average value (μi) for each phoneme and a semivariance matrix (Σi) between various parameters; Reference numeral 5 denotes a word dictionary section in which a word dictionary in which all words to be recognized are expressed in symbol strings in units of phonemes is stored. In the word dictionary, for example, the words ``Satsuporo'' and ``Kankei'' are written as ``SAQPORO'' and ``KAN=NAI.''

Next, the operation of the above conventional example will be explained. The parameter extraction unit 1 analyzes input phonemes every 10 ms frame, extracts parameters, and creates a parameter time series. Next, the probability density calculation section 2 compares the parameters obtained for each frame with the phoneme standard pattern of the phoneme standard pattern section 4, and calculates the probability density of the phoneme. Next, the word recognition unit 3 performs phoneme segmentation for each dictionary item according to the dictionary phoneme series that makes up the dictionary item, and then segments the phoneme according to the type of phoneme and the phoneme corresponding to the phoneme according to the following formula. The likelihood l of the segmented interval is calculated, and the similarity is determined 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
Let Ns and Ne be Ns, and the value of each parameter in the n-th frame be _Co , then the likelihood l _x of phoneme X is defined by the following formula.

φ _i (C _o ) represents the probability density of a certain phoneme i,
It is defined as Eq.

φ _i (C _o )=1/(2π) ^N/2 | Σ _i | ^1/2 exp [
-1/2 (C _o -μ _i ) ^T Σ ^-1 _i (C _o -μ _i )]... _Co : N parameters (vector) in the n-th frame μ _i : Average of parameters of music i Value (vector) Σ _i : Covariance matrix In the formula, the range of i in the denominator summension in probability density division differs depending on the phoneme X. For example, when X is music A, i
The range is five vowels: A, E, I, O, and U. The word similarity L _M obtained above was determined for each dictionary item according to the formula, and the dictionary item with the maximum L _M was selected as a recognized word.

L _M = _NP 〓 ^j=1 l _j /NP... (L _M : Similarity of the M-th word in the dictionary l _j : Likelihood of the j-th phoneme in the dictionary phoneme series NP: Dictionary phoneme class) Figure 2 shows the temporal change in the probability density of each phoneme in the /AN=NA/ part when /KAN=NAI/ (Kannai) is uttered. /AN= in this case
Segmentation and likelihood calculation for the NA/ part are performed using probability density values φ _A , φ N= , φ _N , φ _A for each phoneme /A/, /N ₌ /, /N/, /A/. Segmentation is performed according to time changes. /AN=NA/
In the case of , the segmented interval (a-b) is made to correspond to the first /A/, and according to the formula, l _A is calculated using φ _A , and /N=/, /
For N/ and /A/, l _N= , l _N , and l _A are calculated in the same way.

Figure 3 shows the temporal change in the probability density of each phoneme when the same word /KAN=NAI/ is uttered by different speakers. In Figure 3, the segmentation and likelihood calculation for /AN=NA/ are φ _A ,
This is done by changing φ _N= , φ _N , φ _A over time, but /N
When segmenting =/, the probability density φ _N of the phoneme /N/ that comes after /N=/ is not large enough in the /N/ interval, and φ _N= has a large value in the /N/ interval. and reaches the beginning of the next phoneme /A/. Therefore, the segmentation interval of /N=/ becomes the interval (gh), which includes the interval of /N/, so the segmentation of the next phoneme /N/ after /N=/ is incorrect, and the likelihood l Since _N is also low, words containing two consecutive phonemes, such as a phonic or nasal sound, have the disadvantage of being easily misrecognized.

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

(Structure of the Invention) In order to achieve the above object, the present invention provides repellent sound,
When segmenting and calculating the likelihood of a phoneme sequence with continuous nasal sounds, the accuracy of the segmentation and likelihood calculation can be improved by segmenting and calculating the likelihood of two consecutive phonemes, a nasal sound and a nasal sound. It has the effect of improving.

(Description of Embodiment) An embodiment of the present invention will be described below with reference to FIGS. 1 and 3. In FIG. 1, the phoneme standard pattern is the same as in the conventional example. In a word dictionary, words to be recognized are expressed as phoneme symbol strings. Further, the parameter time series obtained by parameter extraction is the same as in the conventional example. The operation of this embodiment will be explained. First, a parameter extraction section 1 obtains parameters for each frame from input speech, and a probability density calculation section 2 calculates the values of the parameters and probability densities obtained from each phoneme standard pattern. Next, in the word recognition unit 3, segmentation of the phoneme X is performed for each dictionary item in the word dictionary unit 5 according to the dictionary phonetic sequence that constitutes the dictionary item, and the phoneme X and the phoneme The likelihood _lX of the segmented interval is calculated by is dominant until the end of the next nasal sound.
Therefore, the continuous two phonemes of a nasal and a nasal are segmented together, and the likelihood is calculated for the segmented interval. /AN= in Figure 3
Each phoneme in the NA/ part /A/, /N=/, /
Looking at the probability densities φ _A , φ _N= , φ _N , φ _A of N/, /A/, φ _N= has a larger value than φ _N at the /N/ part, and at the beginning of /A/ (h ). Therefore,
Using the value of φ _N= , segment the two consecutive phonemes /N=N/ from g to h, and then set the value of φ _N= for the segmented interval (gh-h). The likelihood l _N=N for two phonemes is calculated using the formula. Here, in contrast to the formula, in the case of ordinary phonemes, the likelihood is calculated using the formula as in the past.

In this implementation, the segmentation and likelihood calculation are performed by combining the phoneme sequences of the pellic and nasal sounds, so there is an advantage that the recognition rate of words containing two consecutive phonemes of the pellic and nasal sounds is improved.

However, the use of symbols is in accordance with the formula.

(Effects of the Invention) As described above, the present invention performs segmentation and likelihood calculation with higher precision than conventional methods by segmenting two continuous phonemes of a nasal and a nasal and performing likelihood calculation. be able to.

[Brief explanation of drawings]

FIG. 1 is a diagram for explaining the word speech recognition method in the conventional method and an embodiment of the present invention, and FIG.
KAN=NAI/ (Kannai) /
Each element /A/, /N of AN=NA/ part
Probability density of =/, /N/, /A/ φ _A , φ _N= , φ _N ,
Figure 3 _shows the temporal changes in φ _A when a _different speaker than in Figure 2 generates / _{KAN =} NAI/. FIG. 1... Parameter extraction section, 2... Probability density calculation section, 3... Word recognition section, 4... Phoneme standard pattern section, 5... Word dictionary section.

Claims (1)

[Claims] 1. Word speech recognition that recognizes words in input speech using a word dictionary that describes the words to be recognized as symbol strings in phoneme units and standard patterns for each phoneme that are expressed as distributions of acoustic parameters for each phoneme. In the method, the input speech is matched with each dictionary entry of a word dictionary,
Segment the input speech for each phoneme according to the dictionary phoneme sequence that constitutes each dictionary entry, and use the standard pattern of that phoneme to calculate the probability density that the segmented speech section is generated from that phoneme. The similarity between each dictionary item and the input speech is calculated using the above probability density value for the segmented speech interval, and when recognizing words, the phonic and nasal sounds in the dictionary word are calculated. A word speech recognition method characterized in that, for a continuous phoneme sequence, two consecutive phonemes, a nasal and a nasal, are segmented together and a likelihood calculation is performed.