JPH0431116B2 - - 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 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. Conventional example first
This will be explained with reference to FIGS. 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 a phoneme standard pattern;
3 is a word recognition unit that performs segmentation for each phoneme, likelihood calculation, word similarity calculation, etc. In addition, 4 represents the distribution of various parameters for each phoneme created in advance through preliminary experiments etc. in the form of the average value for each phoneme (〓 _i ) and the covariance matrix among the various parameters (〓 _i ). A phoneme standard pattern section 5 stores phoneme standard patterns, and a word dictionary section 5 stores a word dictionary in which all words to be recognized are expressed in symbol strings in phoneme units. The word dictionary includes, for example, the words "Satsuporo", "Asahikawa", "Akita",
"Sima", "Sisa", etc. are respectively "SAQPORO",
"ASAHIKAWA", "AKITA", "SIMA",
It is written as "SISA" etc.

Next, the operation of the above conventional example will be explained. The input phoneme is analyzed by the parameter extraction unit 1 every 10 ms frame, parameters are extracted, and a parameter time series is created. The probability density calculation unit 2 compares the parameters obtained for each frame with the phoneme standard pattern, and calculates the probability density of the phoneme generated from the parameter values. Next, in the word recognition unit 3,
Using the above parameters and the obtained probability density value, for each dictionary item, segmentation is performed for each phoneme according to the dictionary phoneme sequence that makes up the dictionary item, and the type of phoneme is determined according to the following formula. Then, the likelihood l of the segmented interval corresponding to that phoneme is calculated, and the degree of similarity is determined as the average of the likelihoods of each phoneme in the dictionary entry.
Here, if the phoneme is X, the frame numbers are N _s and _Ne at the start and end of the segmented section corresponding to X, and the value of each parameter in the nth frame is _Co , then The likelihood l _x of X is defined by the following formula.

φ _i (C _o represents the probability density of a certain phoneme i and is defined as in the equation.

φ _i (C _o )=1/(2π) ^1/2 ｜〓 _i ｜ ^1/2 exp [−1/2(C _o −〓 _i ) ^T _-1 〓 _i (C _o −〓 _i )] …… C: j parameters (vector) in one frame 〓 _i : Average value of parameters of a certain phoneme i (vector) 〓 _i : In the covariance determinant, the range of submension i of the denominator in dividing the probability density is the phoneme It depends on what X is. For example, when X is the phoneme A, the range of i 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 〓 ⁱ⁼¹ l _i /NP …… L _M : Similarity of the Mth word in the dictionary l _i : Likelihood of phoneme i in the dictionary phoneme sequence NP: Number of dictionary phonemes In the above conventional example, , segmentation and likelihood calculation are performed for each phoneme in a dictionary entry using the probability density value of the phoneme. Figure 2 shows the temporal changes in the probability density of each phoneme when uttering /SiMA/ (island). In this case, segmentation and likelihood calculation are performed for each phoneme /S/, /
The probability density values of i/, /M/, /A/ φ _s , φ _i ,
This is done according to the time changes of φ _M and φ _A , and /
Segmentation of S/ is performed using a frame in which φ _s is low and φ _i is high, a is the rear end of /S/, and φ _s is used for the segmented section (SF-a). Perform degree calculations. Similarly, for the second phoneme /i/ following the initial /S/, φ _i
Frame b, where φ _M is low and φ M is high, is taken as the rear end of /M/, and the segmented section (a ~
For b), likelihood calculation was performed using φ _i .

Figure 3 shows the temporal change in the probability density of each phoneme when /SiSA/ (suggestion) is uttered. Segmentation and likelihood calculation are performed for each phoneme /S/, /
The probability density values of i/, /S/, /A/, φ _s , φ _i ,
This is done according to the time changes of φ _s and φ _A , but the initial /
When segmenting S/, φ _i becomes very small because the following /i/ is devoiced, and φ _s is the original section of the word-initial /S/ (FS-c). It also exceeds the original interval (c to d) of /i/ that follows the /S/ at the beginning of the word, so the trailing end e of /S/ that follows /i/ is changed to /S/ at the beginning of the word.
It was output as the rear end of S/, causing a segmentation error.

For this reason, the phonemes /i/, / following the /S/ at the beginning of the word
Segmentation for the S/ and /A/ phonemes is also incorrect and the likelihood becomes low, resulting in the disadvantage that words containing devoiced vowels are easily misrecognized.

(Object of the Invention) The present invention is intended to eliminate the drawbacks of the above-mentioned conventional examples, 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 for each phoneme, and a standard pattern of each phoneme expressed as a distribution of acoustic parameters of each phoneme. When recognizing a word, the input speech is checked against each dictionary item in the word dictionary, and the probability density generated from that phoneme is calculated using the phoneme standard pattern for each phoneme according to the dictionary phoneme sequence that makes up each dictionary item. is calculated, the input speech is segmented, 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 interval. In the speech recognition method, when segmenting and calculating the likelihood of a devoiced vowel sandwiched between unvoiced consonants, the probability density value of each phoneme is used to segment the unvoiced consonant, unvoiced vowel, and unvoiced vowel, including the unvoiced vowel. consonant sequence 3
This method is characterized by segmenting phonemes 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 figure, a parameter extraction unit 1,
The probability density calculation section 2 and the phoneme standard pattern section 4 are the same as those in the conventional example described above, and the differences from the conventional example are as follows.
It mainly includes the contents of the word dictionary section 5 and part of the segmentation and likelihood calculation of the word recognition section 3. The word dictionary stored in the word dictionary section 5 represents the words to be recognized as phoneme symbol strings, but the difference from the conventional example is that vowels that are easily devoiced,
For example, "ASAHKAWA", "AKTA", "S
The I's marked with a circle, such as "MA" and "SSA," are pre-assigned with a code indicating that.

The method of this embodiment first obtains parameters for each frame from the input audio using the parameter extraction unit 1,
Furthermore, the probability density calculation unit 2 uses the values of the parameters to calculate the probability density obtained from each phoneme standard pattern. The process up to this point is the same as the conventional example. Next, in the word recognition section 3, the word dictionary section 5
For each dictionary entry, perform segmentation of phoneme X according to the dictionary phoneme sequence that constitutes that dictionary entry, and calculate the likelihood l _x of the segmented interval corresponding to that phoneme X and that phoneme X. do. When there is a voiceless vowel V sandwiched between voiceless consonants C ₁ and C ₂ in the dictionary phoneme sequence, the probability density value of the voiced vowel does not indicate the nature of the vowel, but rather the nature of the voiceless consonant. Therefore, in the above segmentation, by using the type of each phoneme in the sequence of voiceless consonants, voiceless vowels, and voiceless consonants (C ₁ VC ₂ ), the phoneme, and the corresponding value of each phoneme probability density, The three phonemes are segmented together, and the likelihood l _c1vc2 is calculated for the segmented interval.

In Figure 3, the probability density φ _i of /i/ between /SiS/ is almost nonexistent, and instead, the probability density value φ _s of /S/ at the beginning of the word is at the end of the third S from the beginning of the word.
It is dominant up to e.

Therefore, segmentation is performed using the probability densities φ _S and φ _A of the third phoneme /S/ of the three consecutive phonemes including the devoiced vowel and the vowel /A/ that follows it, and the segmentation result is The likelihood is calculated using φ _s for the interval. By doing this, the successive three phonemes of a voiceless consonant, a voiceless vowel, and a voiceless consonant, /SiS/, correspond to the interval (FS~e) and good segmentation can be achieved, which improves the accuracy of the likelihood calculation. .

In this embodiment, the segmentation likelihood calculation is performed by grouping together the phoneme sequences of voiceless consonants, voiceless vowels, and voiceless consonants, including the voiceless vowel, without treating the voiceless vowel as a single phoneme. This has the advantage of improving the recognition rate for words containing words.

(Effects of the Invention) The present invention uses the probability density value of each phoneme to segment and calculate the likelihood of a devoiced vowel sandwiched between unvoiced consonants. Since the three continuous phonemes of a voiced vowel and a voiceless consonant are segmented and the likelihood calculation is performed, this method has the advantage that the segmentation and likelihood calculation can be performed with higher accuracy than the conventional method.

[Brief explanation of drawings]

FIG. 1 is a functional block diagram of an apparatus for carrying out an example of a conventional word speech recognition method and an embodiment of the word speech recognition method of the present invention, and FIG.
Figure 3 shows the time change in the probability density of each phoneme when uttering SiMA/, (island), /SiSA/
FIG. 4 is a diagram showing changes in the probability density of each phoneme when uttering "(suggestion)". 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. When recognizing words in input speech, the input speech is compared with each dictionary item in a word dictionary in which the word to be recognized is expressed as a symbol string for each phoneme, and the word is expressed in the distribution form of the acoustic parameters of each phoneme. Using the standard pattern of each phoneme, the input speech is segmented by calculating the probability density generated from that phoneme for each phoneme according to the dictionary phoneme series that constitutes each dictionary item, and the segmented speech is When recognizing words by calculating the similarity between each dictionary item and the input speech using the above probability density value for the interval of A word speech recognition method characterized by segmenting consecutive consonants together and performing likelihood calculations using probability density values of phoneme sequences of unvoiced consonants, unvoiced vowels, and unvoiced consonants, including vowels. .