JPS6131878B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a word speech recognition method for recognizing word speech of a limited vocabulary. As a method of word speech recognition, the speech patterns of all the words to be recognized are stored as standard patterns, the similarity between the input speech and each standard pattern is calculated, and the standard pattern word with the maximum similarity is selected. There is a way to use this as a recognition result. However, in this method, it is necessary to store the entire word as a standard pattern, so the storage capacity required to store the standard pattern increases in proportion to the number of vocabulary, and is not suitable when the number of vocabulary is large. For this purpose, a method has been considered in which only the spectral patterns of the phonemes constituting words are stored as standard patterns, and this is used together with a word dictionary that defines the phoneme symbol strings of the words (hereinafter referred to as phoneme sequences). This method first divides speech into relatively short time intervals, calculates the degree of similarity between the spectrum of each division and the spectrum of a standard phonological pattern, and uses the results to create a phonological sequence in a word dictionary. The word recognition unit performs matching, calculates the degree of similarity between the entire input speech and the phoneme sequence, and outputs the word of the phoneme sequence with the maximum degree of similarity as the recognition result. As is well known, the length of the phonemes in a word changes each time it is uttered, even for the same word, and the degree of this change differs depending on the phoneme. There are also large differences in phonological length due to individual differences. Therefore, when calculating the similarity between the input speech and the phoneme sequence, it is necessary to perform matching that allows some variation in the duration of each phoneme. A method that takes this into consideration is ``Transactions of the Institute of Electronics and Communication Engineers Vol. 55-D.
No. 3 (March 1972) Koda et al. "Machine recognition system for digit speech" was proposed. In addition, each phoneme in the dictionary is given a minimum and maximum duration.
A method of reducing the amount of calculation required for matching processing using a dynamic programming method is proposed in ``Japanese Patent Application No. 51-81504.'' However, these methods have the disadvantage that matching with the phoneme sequence must be performed before the end of the speech is known, and they are not suitable for so-called real-time processing, in which matching is performed in parallel with speech input. This invention adds an identification symbol of "1" or "0" to the phoneme symbol of the phoneme series in a word dictionary, and defines a phoneme section that must be associated with a sound.
It shows the distinction between phoneme sections that can be skipped without necessarily being associated with speech, and thereby allows matching to be performed in parallel with speech input, allowing for a certain degree of variation in duration for each phoneme in the phoneme series. The aim is to realize a real-time word speech recognition method that uses phonemes as recognition units. FIG. 1 is a block diagram showing an embodiment of the word speech recognition method according to the present invention. The audio input from the input terminal 1 is divided into relatively short time intervals (referred to as frames), for example, about 15 msec, by a feature extraction unit 2 that extracts audio feature parameters.
Feature parameters expressing the power and spectrum are calculated for each of the sections. Although various feature parameters can be considered to express the spectrum, an autocorrelation function of speech is used here as an example. The phoneme standard pattern storage unit 3 is a part that stores the phoneme spectrum necessary for recognition as a standard pattern, and the degree of similarity between the phoneme standard pattern and the spectrum of each segment of the input speech is stored in the phoneme recognition unit 4.
Calculated in . Various measures can be considered to express the degree of similarity between the spectra, such as the Euclidean distance between feature parameters and the integration of the square of the difference in power spectra over the entire band. Here, as an example, a case will be described in which log likelihood (hereinafter simply referred to as likelihood) based on the maximum likelihood spectrum estimation method is used as the degree of similarity. According to the maximum likelihood spectrum estimation method, the autocorrelation function of the n-th segment of the input speech is expressed as {v ₀ ⁽ⁿ⁾ , v ₁ ⁽ⁿ⁾ , v _2(o) ... v _p ⁽ⁿ⁾ } (p is the maximum delay time ) (1), and the maximum likelihood spectral parameters of the standard pattern of phoneme /x/ are {A ₀ ^(x) , A ₁ ^(x) , A ₂ ^(x) ...A _p ^(x) } (2) The likelihood l(n, x) indicating that the n-th division of the input speech is likely to be the phoneme /x/ is calculated by the following equation. however

Constant terms that depend only on [Formula] are excluded. Furthermore, the maximum likelihood spectral parameters expressing the standard pattern of phoneme

[Formula] can be easily calculated in advance from the autocorrelation function of the speech corresponding to the phoneme, so it is better to store it in the phoneme standard pattern storage unit 3 in the form of maximum likelihood spectral parameters than the autocorrelation function (3 ) is convenient for calculating the likelihood of the equation. The likelihood based on equation (3) is calculated for all phonemes for each speech segment and stored in the likelihood storage unit 5. A word dictionary storage unit 6 stores the classification series of the input speech and the phonological series of the word to be recognized based on this likelihood.
The matching part 7 performs matching with the phonological series.
It will be held in The principle and operation will be explained below. _The length of _the input speech _is divided into N parts, _one of the phoneme sequences to be matched with _the speech _is set as Assume that there are restrictions on the maximum duration D j and the maximum duration D _j . Matching the input speech with this phoneme sequence to find the correspondence relationship is the point at which (J-1) phonemes change in the input speech.

This is equivalent to determining [Formula] so that it satisfies the following conditions. Under these conditions, the degree of similarity L(x) between the input speech and the phoneme sequence is defined as follows. That is, the likelihood of the standard pattern of phonemes to be associated with each segment of the input speech is determined and added for all segments. The likelihood is added at the point of phonological change.

There are as many combinations as [Formula], and the largest one is selected as the similarity. In this invention, the following steps are taken to efficiently calculate this degree of similarity. The length of the same phoneme symbols of each frame time length arranged by the maximum duration of each phoneme in the phoneme series X

[Formula] Phonological series , and call it the phonological sequence X′. For each phoneme symbol of this phoneme sequence X', identification symbols of "1" are attached from the beginning for the minimum duration of the phoneme, and identification symbols of "0" are attached to the remaining symbols. The minimum duration d ₁ of the first phoneme of phoneme series X is 1
If the above is the case, there is only _one correspondence between the beginning of the input speech and the phonetic sequence There are as many as the maximum duration D ₁ +1. If the minimum duration d ₂ of the second phoneme is also 0, it increases further. For this reason, it is necessary to check the identification symbol of each phoneme symbol of the phoneme sequence X' from the beginning before matching it with the input speech. As a result, if the identification symbol of the _m0th phonetic symbol is "1" for the first time, there are _m0 possible associations between this phonetic sequence X' and the beginning of the input speech. Therefore, the similarity between the partial sum of input speech up to the n-th segment and the m-th subsequence of the phoneme sequence X′ is S
When expressed as (n, m), the initial condition S ( ₀ , m) = 0, (m = 0, 1, 2... m ₀ ^-1 )
(7a) S( ₀ , m)=C, (m=m ₀ , m ₀ ⁺¹ ...M) (7b) S(n, ₀ )=C, (n=1, 2...N) (7c ) (where C is a very large negative constant), the following recurrence formula (i) The m-th phoneme of the phoneme sequence X′ is /x _i / and its identification symbol is “1 ” then S (n, m) = S (n _-1 , m _-1 ) + l (n, x _i )
( _{8a )} (ii) When the m _- th phoneme of the phoneme sequence +l(n, x _i ), S(n, m _-1 )} (8b) By calculating the similarity L(x) in equation (6) in the range of 1≦n≦N, 1≦m≦M is obtained as S(N,M). An example of the procedure described above will be explained. Assume that one phoneme /x _i / of the phoneme sequence X is restricted to a minimum duration of 2 and a maximum duration of 5. It is expressed as the following sequence in the phonetic sequence X′. The identification symbol is shown here. X'：...x _i-1 ⁽⁰⁾ x _i (1)x _i (1)x _i ⁽⁰⁾ X _i ⁽⁰⁾ x _i ⁽⁰⁾
x
_i+1 (1)... Figure 2 illustrates part of the calculations for this phoneme sequence X'. _The horizontal axis corresponds to the segment number of ^the phonetic _{sequence ,} and the vertical _axis corresponds to the segment number of the phonetic sequence X _{' .} ⁰⁾ respectively. In Figure 2, if we assume that each point on the n _-1th column has already calculated the degree of similarity when matching the voices up to the n _-1 category, then the next point is the voice in the nth category. phoneme/x
When the likelihood l(n, x _i ) for _i / is given, the similarity up to the n-th division is calculated as follows. First, point ₂₀₀ has similarity S( _{n -1 ,}
m _-5 ) has been found, add l(n, x _i ) to this, and calculate the similarity S when phoneme /x _i / is one category.
(n, m ₋₄ ) is calculated at point 206. Also 20 points
1 is the similarity S when the phoneme /x _i / is in one category.
Since (n _-1 , m _-4 ) is found, similarly l
(n, x _i ) is added to calculate the similarity S(n, m ₋₃ ) at point 207 when the phoneme /x _i / is divided into two categories.
At point 202, the similarity S (n _-1 , m _-3 ) when the phoneme /x _i / is divided into two categories is found, so the phoneme /x _i /
When there are 3 categories, the similarity is S (n _-1 , m _-3 ) + l
(n, x _i ), but in this case, the phoneme/
Since the identification symbol of x _i / is "0", it is compared with the similarity S (n, m _-3 ) in the case of two classifications, and the larger value is set as S (n, m _-2 ) and the point 208 assigned to. That is, at point 208 there is a phoneme /x _i /
The degree of similarity is assigned when there are two or three categories. By similar means, point 209 is further assigned a degree of similarity when the phoneme /x _i / is in 2 to 4 categories, and point 210 is assigned a degree of similarity when it is in 2 to 5 categories. Although the above is only an example of calculation for a part of the phoneme sequence X', it shows that the results from up to n _-1 divisions to n divisions can be easily calculated. Therefore, by calculating equation (8) from the first section at the beginning of the input speech to the Nth section at the end, the final result of the recurrence formula S(N, M) is calculated by calculating the degree of similarity between the input speech and the phoneme sequence X. L(X) is obtained. FIG. 3 shows an example of matching with the phoneme sequence "7" /nana/. Note that equation (8) is

It is only necessary to calculate within the range of [Formula], so for each phoneme series in the word dictionary, the length

By retaining the intermediate results of [Formula], it is possible to match multiple phoneme sequences in parallel with speech input. The part that executes the calculations described above is the matching unit 7 shown in FIG. The addition S (n _-1 , m _-1 ) + l (n, x _i ) of equations (8a) and (8b) is performed in the adder 71, and the comparison for the max operation of equation (8b) is performed using the comparator. 72, and its output is temporarily stored in a register 73 so that the max calculation of equation (8b) can be performed efficiently. The intermediate result similarity for each phoneme sequence is stored in a similarity calculation buffer 74. First, prior to matching, initial values of equations (7a) and (7b) are set in the similarity calculation buffer 74. n
The operation for the audio segment is performed as follows. An address designation signal is sent from the control section 8 to the dictionary storage section 6, and the symbol and identification symbol representing the phoneme x _i of the phoneme series X' to be matched are read out. The number representing x _i further serves as an addressing signal for the likelihood storage section 5 and is used to read out the likelihood l(n, x _i ) for the phoneme x _i from the likelihood storage section. The identification symbol becomes a control signal for the comparator 72. On the other hand, from the similarity calculation buffer 74
One of _{the matching results S(n -1 ,}
m _-1 ) is read. In the adder 71, this S
(n _-1 , m _-1 ) and the likelihood l(n, x _i ) read from the likelihood storage unit 5 and the resulting value S(n _-1 , m _-1 ) + l (n, x _i ) is sent to the comparator 72. In the comparator 72, if the identification symbol read out at the same time as the phoneme x _i is "1", S (n _-1 , m _-1 ) + l
(n, x _i ) is output as is, and the identification symbol is “0”
If so, it compares with the previous calculation result S(n, m _-1 ) stored in the register 73 and outputs the larger value. This output is the interim result of similarity S(n,
m) is stored in the similarity buffer 74 and simultaneously stored in the register 73, and used in the next calculation. Note that each time the phoneme series to be matched changes, it is necessary to first input a large negative constant C into the register 73 as the initial value of equation (7c). The above processing is performed as many times as necessary for each phoneme sequence.

Repeat [Formula] and continue until the end of the voice N. Then, the similarity S(N, M) obtained for each phoneme sequence is stored in the similarity storage section 9, and the matching is completed. Next, the operation of the control section 8 will be explained. The control unit 8 detects the start and end of audio, controls the matching unit 7,
Controls the maximum similarity value selection unit 10, etc. That is, first, a power value is received for each section from the feature extractor 2, and the starting end is detected depending on whether the power exceeds a predetermined threshold. When the starting edge is detected, an initial value is set in the similarity calculation buffer 74 and matching is started. Furthermore, the power of each segment is checked, and when the power becomes smaller than the threshold value, the previous segment is selected as a candidate for the end of the voice. The similarity of the matching results up to the final candidate category is sent to the similarity storage section 9, and the maximum similarity selection section 10 is operated. Maximum similarity selection unit 10
Then, the one with the highest degree of similarity to each phoneme sequence is selected, and the word in that phoneme sequence is used as a candidate for the recognition result. After that, if a segment in which the power is smaller than the threshold continues for a certain period of time or more, that candidate word is output to 11 as the final recognition result. If the power increases again within a certain period of time, it is determined that this is a silent section within the word, and detection of the end of the voice is continued. As described above, according to the present invention, by adding an identification symbol of "1" or "0" to each phoneme in the phoneme series prepared as a word dictionary, matching can be performed with a limit on the duration of the phoneme. A word speech recognition device that can be expected to have a high recognition rate and can perform real-time processing can be realized. Note 1
The log likelihood obtained by the maximum likelihood spectrum estimation method used to express the degree of similarity between the speech of the segment and the standard pattern of phoneme is one example, and the present invention is not limited thereto.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the word speech recognition method according to the present invention, FIG. 2 is a diagram for explaining the operation of a matching section, and FIG. 3 is a diagram showing an example of matching. 1: Audio signal input terminal, 2: Feature extraction section, 3:
Phonological recognition unit, 4: Standard pattern storage unit, 5: Likelihood storage unit, 6: Word dictionary storage unit, 7: Matching unit, 71: Adder, 72: Comparator, 73: Register, 74: Similarity calculation buffer , 8: control section,
9: similarity storage unit, 10: similarity maximum value selection unit,
11: Recognition result output terminal.

Claims (1)

[Claims] 1. A feature extraction unit that divides input speech into relatively short fixed time intervals and calculates an autocorrelation function representing the power and spectrum of each division, and converts some phoneme spectral patterns into standard patterns. A phonological recognition unit that calculates the degree of similarity between the autocorrelation function for each category and the standard pattern; The similarity between the input speech and the phoneme symbol sequence is calculated based on a fixed-form recurrence formula while referring to the word dictionary memory that stores the symbol sequence as a word dictionary and the similarity calculated by the phoneme recognition unit. A word speech recognition device consisting of a matching section and a maximum similarity selection section that selects the maximum similarity among the similarities obtained by the matching section and outputs the word of the corresponding phonetic symbol sequence as a recognition result. In the word dictionary memory, for each phoneme constituting a word, the same phoneme symbols are arranged at the fixed time interval for the maximum duration length, and identification symbols of "1" are stored for the minimum duration length among these phoneme symbols. , and the remaining phonetic symbol sequences with identification symbols of "0" are stored, and the matching section combines the n-th segment of the input speech of length N with the m-th phonetic symbol of the length M phonetic symbol sequence. In the recurrence formula calculation for, if the identification symbol attached to the phonetic symbol is "1", the phonetic symbol
The similarity l(n, x) between the standard pattern of x/ and the input voice of the nth division is the calculation result of the (n-1)th division.
The value added to S (n-1, m-1) is the calculation result S (n, m) for the nth division, and when the above identification symbol is "0", the above similarity l (n, x ) as the calculation result S (n-1, m-
1) and the calculation result S (n,
m-1), the larger value is calculated as the calculation result S (n, m) of the n-th division, and the similarity between the input speech and the phonetic symbol sequence is calculated as the calculation result of the N-th division.
A word speech recognition method characterized by calculating by S (N, M).