JPS60202496A - Word voice recognition - 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.

(Structure of conventional example and its problems) A conventional word speech recognition method will be explained with reference to FIGS. 1, 2, and 3. In FIG. 1, the word dictionary of the word dictionary section 3 represents all the words to be recognized in phoneme sequences. For example, the words "saraporo" and "fussa" are written as r5AQ.
POROJ r ) IUQSA J etc.

A parameter extraction unit 1 analyzes input audio every frame of 10m5, extracts parameters, and creates a parameter time series. Next, the word recognition unit 2 performs phoneme segmentation for each phoneme according to the dictionary phoneme series that constitutes the dictionary item for each dictionary entry using the above-mentioned parameter. , the likelihood t of the segmented interval corresponding to that phoneme is calculated using the above-mentioned y parameter, and the degree of similarity is calculated as the average of the likelihoods of each phoneme in the dictionary entry according to formula (2).

The word similarity obtained above is determined for each dictionary item, and the dictionary item with the maximum LM is determined as a recognized word.

LM=ΣLi/NP... ■ i=1 In the above method, segmentation and likelihood calculation of consonants are performed using the logarithmic normalized power of the speech shown in the formula (■), P (N is the frame number (flat) )) and the value of the inter-adjacent frame cepnutrum distance CDH shown in Equation 0. Segmentation is performed based on the value of the cepnutrum distance CDH between adjacent frames, and the likelihood of the consonant is calculated from the segmented duration length of the consonant and LNG according to Equation 0.

FIG. 2 shows the temporal change in the logarithmically normalized A'ku (N) of the voice when /5AQPORO/ (Sapporo) is uttered.

In this case, the segmentation and likelihood calculation of the consonant /Ql are as follows: The size of P(6) from the rear end frame a of /V is T
Search for frames below P and check whether PH is above TP or the quezonutrum distance CD (transition) between adjacent frames.
is the threshold value T. For the frame (b) that is larger than D, the likelihood is calculated according to the formula 0 using the duration length LNG of the segmented section (a-b) of the consonant /Q/, which ends the frame at the end of the consonant /Q/.

FIG. 3 shows the temporal changes in the voice power P (goods) and the inter-adjacent frame casistrum distance CDfi when /HTJQ8A/ (Fussa) is uttered. Figure 3 consonant/Q
Focusing on the / part, when searching from the trailing frame (C) of /U/ to the trailing frame (d) of A/, the size of P@ is smaller than that of /Q/ in Figure 2. In the Q interval,
It never goes below the threshold Tp, and the value of CDl[ is also T.

It will never be higher than D. Therefore, when performing segmentation of the consonant Q, it passes through the rear end (d) of the original 7v interval and searches for the rear end frame (d) incorrectly, lowering the likelihood that the consonant and voiceless fricative are consecutive. There was a drawback that words containing phoneme sequences were incorrectly recognized (5).

(Objective of the Invention) An object of the present invention is to eliminate the drawbacks of the prior art described above, improve the accuracy of segmentation and likelihood calculation, and thereby improve the word recognition rate.

(Structure of the Invention) In order to achieve the above object, the present invention, when performing segmentation and likelihood calculation of a phoneme sequence in which a consonant and a voiceless fricative are continuous, segments two consecutive phonemes of a consonant and a voiceless fricative together, Next, using the standard no-e tan expressed in the distribution form of the acoustic parameters of phonemes, calculate the probability density that the segmented speech section is generated from each phoneme, and calculate the probability density for the segmented speech section. The likelihood calculation is performed using the above probability density value.

(Description of Examples) Examples of the present invention will be described below with reference to FIGS. 3 and 4. FIG. 4 is a functional block diagram of a device for executing the method of this embodiment (6). Grammeter extraction unit 1, phoneme probability density calculation unit 2, word recognition unit 3
, a phoneme standard button part 6, a word dictionary 7, and the like. What differs from the conventional example shown in FIG. 1 is that it includes standard phonemes represented by acoustic parameter distributions. Furthermore, in the word dictionary, words to be recognized are expressed as phoneme symbol strings, and codes are attached in advance to the two consecutive phoneme sequences of consonants and voiceless fricatives to identify them.・
The /P parameter time series obtained by mother meter 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 the input speech, and then, using the value of the four parameters, a probability density calculation section 2 calculates the probability density obtained from the standard bang of each phoneme. Next, the word recognition unit 3 performs the following for each dictionary item:
Segmentation of phoneme X is performed according to the dictionary phoneme series that constitutes the dictionary entry, and the phoneme X and its phoneme
We calculate the likelihood tX of the segmented interval corresponding to The nature of the part is different from shi, and the nature of the part of the consonant is close to that of the voiceless fricative. Therefore, using the probability density value of the unvoiced fricative, two consecutive phonemes of the consonant and the unvoiced fricative are collectively segmented and the likelihood is calculated.

Figure 3 shows the sound when /HTJQSA/ is uttered, the voice P (he), the Kenonutrum distance between adjacent frames, and the CDI.
J) and phoneme /%', A/, /S/, '/
The probability density of IV' is φ□.

φ0. φ8. It shows the change in φe over time. In Figure 3,
The power P of the part of the consonant A/ does not fall below the threshold TP, and is about the same as the power of the next phoneme A/, and the cepstral distance between adjacent frames at the boundary (d) with /V, CD
The value of (d) is also the threshold value T. It does not exceed D and there is no significant change.

Also, the probability density of the /Q/ section is the probability density of /S/, φ8 is dominant, and is dominant up to the rear end (f) of /S/. Therefore, for the 8 series of two consecutive phoneme systems of consonants and voiceless fricatives, we regard the two consecutive phoneme series as voiceless fricatives with a long duration, and use the probability density of voiceless consonants to calculate the rear end of /S/ (f
), perform segmentation, and find the segmentation interval length, LQ8, and the probability density value φ of the voiceless fricative.
8 is used to calculate the likelihood tQ8 of two consecutive phoneme sequences of consonants and voiceless fricatives according to formula (■).

In this example, two consecutive phoneme sequences of a consonant and a voiceless fricative are regarded as one unvoiced fricative with a long duration, and segmentation and likelihood calculation are performed using the probability density of the unvoiced fricative. This has the advantage of improving the recognition rate for words containing continuous phoneme sequences.

C, D: Constant (9) (TLGQ, TL, TH are prepared in advance through preliminary experiments, etc.) φ, (CN) represent the probability density of a certain phoneme i,
It is defined as equation 0.

CN = 4 parameters (vector) μ in the Nth frame: Average value (vector) of the parameters of a certain phoneme i Σ: Covariance matrix The range is 15 phoneme groups, including 5 vowels, nasals, voiced consonants, and voiceless consonants.

(Effects of the Invention) When performing segmentation and likelihood calculation of a phoneme sequence consisting of two consecutive consonants and unvoiced fricatives, the present invention uses the probability density of unvoiced fricatives to segment two consecutive phonemes of consonants and fricatives together. By performing the likelihood calculation, there is an advantage that segmentation and likelihood calculation can be performed with higher precision than the conventional method.

[Brief explanation of the drawing]

Figure 1 is a diagram for explaining the word speech recognition method in the conventional example, and Figure 2 shows the normalized logarithmic power P■ of the voice when uttering /5AQPORO/ (Sapporo), the cepstral distance between adjacent frames, CD- Figure 3 shows the time change of /) IUQSA/ (Fussa).
(N), CD- and the phoneme /) I/ , /U/ , /
S/. The probability density of A/φ,,φ. , φ8. φ. FIG. 4 is a diagram for explaining a word speech recognition method in an embodiment of the present invention. 11... Rigram meter extraction section, 12... Phoneme probability density calculation section, 13... Word recognition section, 14... Phoneme standard slam section, 15... Word dictionary section. (11) Figure 1

Claims (1)

[Claims] The input speech is compared with the dictionary entries of a word dictionary in which words to be recognized are expressed as symbol strings in phoneme units, and the input speech is divided into units for each dictionary according to the dictionary phoneme series that constitutes each dictionary entry. Segmentation is performed on the input audio using the acoustic parameter analyzed for each time, and the above acoustic i+ parameter is used for the segmented section,
In the word speech recognition method, which recognizes words by determining the similarity between each dictionary item and the input speech, for a phoneme series in which a consonant and a voiceless fricative are consecutive in a dictionary word, two consecutive phonemes of a consonant and a voiceless fricative are grouped together. The segmented speech (1) is segmented by segmentation, and the probability density that the segmented speech section is generated from each phoneme is calculated using a standard baton expressed in the distribution form of the acoustic parameters of the phoneme. A word speech recognition method characterized in that likelihood calculation is performed using the above probability density value for the interval.