JPS6310840B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a phoneme identification method in a speech recognition device.

Conventional examples and their problems The phoneme identification method in conventional speech recognition devices is
First, the time-series pattern of the feature vector of the input speech is roughly classified into phoneme intervals, and after separating, for example, non-stationary consonant intervals and relatively constant vowel intervals, each frame of the consonant interval is classified. The consonant was determined by calculating the distance between all consonants and the standard pattern, and the same calculation was performed for the vowel interval to determine the vowel. Conventional speech recognition devices, such as those mentioned above, use a rough classification method for phoneme intervals that involves simply classifying consonant and vowel intervals based on sequence correlation values, etc., and the final decision on phoneme is made by comparison with a standard pattern. There are things that are going on. However, in this case, the accuracy of the rough classification itself is poor, and the set of standard patterns is also large, resulting in an enormous amount of calculation for pattern matching. Therefore, in order to improve these shortcomings, the consonant intervals can be further classified into voiced consonant intervals and unvoiced consonant intervals, and further improve the precision by using information regarding the bias toward low or high frequencies in the rough classification of phoneme intervals. Among those that perform high classification, there are also those that classify voiceless consonant intervals into plosives and fricatives. Then, by reducing the set size of the standard patterns for which final decisions are to be made, the number of comparisons with the standard patterns for which final phoneme decisions are to be made is reduced, thereby improving processing speed. However, even in this case, although the comparison circuit with the standard pattern is reduced, the amount of calculation required for rough classification increases, and when evaluating the device as a whole, the amount of calculation is not necessarily reduced, and the amount of calculation required for phoneme identification increases. There is a problem with calculation time.

Purpose of the Invention The present invention solves the above-mentioned conventional drawbacks by simplifying the distance calculation between the standard pattern and the feature vector of input speech, reducing the processing time required for phoneme identification, and enabling real-time processing by speech recognition devices. The purpose is to simplify the device configuration.

Structure of the Invention In order to achieve the above object, the phoneme identification method of the present invention includes a feature sequence conversion means for converting input speech into a time series pattern {〓 _t1 , 〓 _t2 , ... 〓 _tN } of a feature vector {〓 ti } _. , a standard pattern storage means for storing a standard pattern of feature vectors of each phoneme, and a distance determination means for calculating the distance between the input and the feature vector of the standard pattern, and the distance determination means is configured to be a standard pattern storage means for storing a standard pattern of feature vectors of each phoneme. In the determined frame {〓 _tk }, the phoneme is determined by comparing with all standard patterns, and for subsequent frames {〓 _t 〓} (however, τ>k), the determined standard pattern is used. If the distance is less than a predetermined threshold, the frame of the feature vector to be compared as the same phoneme is updated, and if it is more than the threshold, a new phoneme is created by comparing with all other standard patterns. and the subsequent frames {〓 _tn }
(However, for m>τ), the configuration is such that phoneme identification is performed by repeating determination using a similar distance threshold.

DESCRIPTION OF EMBODIMENTS Hereinafter, one embodiment of the present invention will be described based on the drawings.

The processing concept of the present invention will be explained using FIG.

The phoneme identification means includes a feature system conversion sequence means 1 that converts an input speech time series into a feature vector sequence {〓 _ti } at regular time intervals, and a rough phoneme (voiceless consonant, voiced consonant, vowel, etc.) from the feature vector sequence. a major classification means 2 for classifying a voiced consonant and a vowel; a distance determining means 3 for determining a phoneme by calculating the distance between a voiced consonant and a standard pattern 4, 4' of a vowel in the case of a voiced consonant and a vowel;
3', and a phoneme sequence output means 5 for outputting a phoneme time sequence.

The operation of FIG. 1 will be explained. Input speech is converted into a sequence of feature vectors by the feature sequence conversion means 1. The feature vectors are roughly classified into voiceless consonants, voiced consonants, and vowels by the rough classification means 2, for example, based on their bias toward low or high frequencies. In the case of voiced consonants and vowels, distance calculations between voiced consonants and vowel standard patterns 4 and 4' are performed by distance determinations 3 and 3', respectively;
Assign the closest phoneme. When each phoneme is determined, the phoneme sequence output means 5 outputs the phonemes in time series. The reason why voiceless consonants are not compared with the standard pattern in the above is due to the non-stationarity of voiceless consonants.

Next, the detailed configuration and processing of the distance determining means 3 in the present invention will be explained using FIGS. 2 and 3.

The distance determining means 3 includes a feature vector storage unit 6 that stores input feature vectors, and one standard pattern vector 4 of the feature vector and the standard pattern 4.
The distance calculation unit 7 that calculates the distance between the first and first determined feature vectors compares all standard pattern vectors, and once the phoneme is determined, the next input feature vector is the first determined feature vector. It consists of a comparison control unit 8 that compares standard pattern vectors, and if the distance is less than a certain value, it is considered the same phoneme, and if it is more than a certain value, it compares with all other standard pattern vectors and updates the phoneme. .

FIG. 3 shows the detailed operation of the comparison control section 8.

In Fig. 3, the feature vector is two-dimensional (X _ti1 ,
X _ti2 ). The standard pattern vectors of phonemes |A|, |i|, |u| have their centroids represented by 9, 10, and 11. The input feature vector is represented by 12 as a sample of time point (t) with {〓 _tk } as the starting point. Also, the constant threshold values corresponding to each standard pattern vector are r ₁ ,
It is expressed as r ₂ and r ₃ . The comparison control unit 8 calculates the distance between the feature vector {〓 _tk } at a certain time and all the standard pattern vectors 9, 10, and 11.
In FIG. 3, the vowel |i| is determined. The next feature vector {〓 _tk+1 }' is determined by calculating only the distance of |i| from the standard pattern vector 10 and comparing it with a certain threshold value r _2. In this case, the following is obtained, so the vowel |i| shall be. The same applies to the feature vector {〓 _tk+2 }. Since the feature vector {〓 _tk+3 } is equal to or greater than a certain threshold value r ₂ , the distance from the other standard pattern vectors 9 and 10 is calculated, and in this case it is determined to be |A|. Regarding the feature vector {〓 _tk+4 }, the distance from the standard pattern vector 9 of |A| is r ₂ or more, so the other standard pattern vector 1
0 and 11 are also compared, and in this case as well, it is determined to be |A|. Since the feature vector {〓 _tk+5 } {〓 _tk+6 } {〓 _tk+7 } has a distance of r ₁ or less from |A|, all of them are determined to be |A|. In the case of FIG. 3, the number of distance calculations required for phoneme determination not according to the present invention is 8×3=24, and the number of distance calculations according to the present invention is 5+3×3=
This is 14 times, which is almost half the number of calculations. Since stationarity is large in vowel intervals, this method can be expected to reduce the number of calculations even more.

Effects of the Invention As explained above, according to the present invention, the following effects can be obtained.

The present invention determines pattern matching between a feature vector of an input voice and a standard pattern by comparing it with all standard patterns in a predetermined frame, and in subsequent frames of the feature vector,
Only the distance to the determined phoneme is calculated, and if it is less than a certain threshold, it is considered the same phoneme, and if it is more than a certain threshold, it is compared with all other standard patterns, the phoneme is updated, and the same operation is performed. Since phoneme identification is performed by repeatedly applying , the processing time required for pattern matching for phoneme determination can be shortened, and the processing configuration of the speech recognition device can be simplified.

[Brief explanation of the drawing]

FIG. 1 is a specific configuration diagram of the phoneme identification means for calculating a phoneme sequence from input speech according to the present invention, and FIG. 2 is a distance judgment means for calculating the distance between a feature vector and a standard pattern in the phoneme identification means according to the invention. FIG. 3 is an explanatory diagram of the operating principle of the distance determining means in the present invention. 1...Feature series conversion means, 3...Distance determination means, 4...Standard pattern.

Claims (1)

[Claims] 1. A feature sequence conversion means for converting input speech into a time series pattern of feature vectors {〓 _ti } {〓 _t1 , 〓 _t2 , ...〓 _tN }, and a standard pattern storage means for storing a standard pattern of feature vectors for each phoneme. and a distance determining means for calculating the distance between the feature vectors of the input and standard patterns, and the distance determining means is configured to calculate the distance between all standard patterns _and For subsequent frames {〓 _t 〓} (where τ > k), if the distance from the determined standard pattern is less than a predetermined threshold, the phoneme is determined by comparing The frame of the feature vector to be compared is updated, and if it is equal to or higher than the threshold, a new phoneme is determined by comparison with all other standard patterns, and subsequent frames {〓 _tn } (where m> For τ), this phoneme identification method is configured to perform phoneme identification by repeating determination using a similar distance threshold.