JPS6048040B2 - Learning processing method for individual differences in speech recognition - Google Patents

Abstract

PURPOSE:To attain a high aural confirmation rate and thus to increase the using efficiency be drawing out the standard pattern of each phoneme from part of the study sample. CONSTITUTION:When the new speaker unused for the calculation of the conversion formula pronounces part of the study words to be recognized, the average corelation coefficient is calculated between various phonemes and their words based on the aural waves. The value obtained through the above calculation is then converted into the logarithm cross-sectional area ratio. Using the value of the area ratio and the conversion formula, the value equivalent to the logarithm cross-sectional area ratioconverted from the averaged correlation coefficient for all words to be recognized is calculated for both the phoneme contained in the study word and the phoneme which is not contained. In this way, the standard pattern based on each phoneme is produced in accordance with the individual difference.

Description

[Detailed Description of the Invention] The present invention provides a method for learning and processing individual differences in speech recognition,
In particular, the present invention relates to a learning method for speech recognition in which a standard pattern adapted to individual differences in speech waves is efficiently created from a small number of learning words (learning samples) in a speech machine recognition device.

Generally, when mechanically recognizing the meaning of uttered speech, the input speech is expressed in an appropriate format, the similarity between the speech and a pre-stored standard pattern is measured, and the The content of the input voice is determined according to the degree of the input voice. In addition, a frequency spectrum or something similar is generally used to express input speech and standard patterns, but even if the same content or words are uttered, the frequency spectrum may differ depending on the speaker. Therefore, if the similarity with the input speech is measured using a standard pattern common to all speakers, it is not possible to recognize the content of the input speech with high accuracy. Therefore, in conventional methods, each time the speaker changes, the standard pattern is recreated by having the speaker speak all the words to be recognized as a predetermined learning sample, or by re-creating a standard pattern by having the speaker speak all the words to be recognized as a predetermined learning sample. A method has been used in which standard patterns of phonemes that are not included in the training samples are estimated using multidimensional conversion formulas by asking people to speak. However, in these methods, as the range of speech content to be recognized expands to, for example, several hundred words, the number of standard patterns increases, and it is necessary to utter all of the words to create a standard pattern. It takes a lot of effort and time to fix it,
The method of estimating the standard pattern of phonemes that are not included in the uttered speech (learning sample) by uttering a certain part of the speech cannot remove the influence of the utterance being biased against the whole recognition target. Therefore, it is not possible to recognize the content of input speech with high accuracy. This drawback poses a very large restriction in making a speech recognition device practical. In order to eliminate the drawbacks of the conventional example, the present invention temporarily creates a standard pattern of phonemes included in some learning samples of the recognition target, and based on this value, For both the phonemes included in the learning sample and the phonemes not included, calculate the values corresponding to the standard pattern obtained when all the words to be recognized are used as the learning sample, and calculate the standard pattern for each phoneme. The purpose of this study is to provide an improved learning method for speech recognition.

Embodiments will be described in detail below with reference to the drawings. The figure shows an embodiment of the present invention, where 1 indicates the sound 4. voice signal input terminal; 2 is a low-pass filter; 3 is an analog-to-digital conversion circuit; 4 is an autocorrelation coefficient calculation circuit; 5 is a phoneme segmentation circuit; and 6 is an intra-segment autocorrelation coefficient averaging circuit. 7 is a switch, 8a and 8b are terminals of switch 7, and 9 is an autocorrelation coefficient all-word averaging circuit that averages the autocorrelation coefficients in the same category. 10 is an average correlation coefficient storage unit, 11 is a log cross section ratio calculation circuit, 12 is an all word learning log cross section ratio storage unit, and 13 is an auto correlation coefficient 14. A partial word averaging circuit that averages self-correlation coefficients corresponding to the same phoneme extracted from a predetermined number of words, that is, partial words;
15 is an average correlation coefficient storage unit, 15 is a logarithmic cross section ratio calculation circuit,
16 is a partial word learning log cross section ratio storage unit, 17 is an autocorrelation coefficient averaging circuit that averages the autocorrelation/number for the same phoneme extracted from all speakers, and 18 is a logarithm. 20 is a switch, 21a, 21b are terminals of switch 20, 22 is a normal matrix calculation circuit, 23 is a normal equation calculation circuit, 24 is a conversion formula storage unit, 25 is a standard pattern calculation circuit, 26 is a maximum likelihood spectral parameter calculation circuit, 27 is a standard pattern storage section, and 28 is a weight storage section.

Next, the operation of this embodiment will be explained.

First, the switch 20 is connected to the terminal 21a, and the standard pattern storage section 27 stores appropriate values calculated empirically or by analyzing the voice of a particular speaker as standard patterns for each phoneme. . Therefore, all words to be recognized are uttered in a predetermined order as learning samples. This analog waveform is input to the audio signal input terminal 1. This analog audio is processed by a low-pass filter to remove high frequency components of, for example, 3.4 KHz or higher, and then converted to an analog to digital converter circuit 3 for example, 8 KHz11.
After being sampled and quantized with 1 bit and converted into a digital waveform, the autocorrelation coefficient calculation circuit 4 generates a
A section of a certain length of time is cut out at a cycle of 5ms, etc.
For example, autocorrelation coefficients up to a finite dimension up to 1 European are calculated. The tth time point within the cut out section (1 t<N
．． N is the number of sample points corresponding to the length of the cutout) When the audio waveform is expressed as Xt, the τ-order autocorrelation coefficient is defined by the following equation. The λT-phoneme section segmentation circuit 5 calculates the likelihood of each phoneme using this autocorrelation coefficient and the standard pattern of each phoneme stored in the standard pattern storage section 27, and uses the value as Accordingly, the uttered speech is segmented into each phoneme interval.

The standard pattern (maximum likelihood spectral parameter) of phoneme j is Aj (7) (a) × τ × τ. Nax)
The likelihood of phoneme j is defined by the following formula. τ-1 The correlation coefficients of each phoneme section segmented in this manner are sent to the next intra-section autocorrelation coefficient averaging circuit 6, where an average correlation coefficient is calculated for each phoneme section.

The switch 7 is initially connected to the terminal 8a, and the average correlation coefficient of each phoneme section is determined by the autocorrelation coefficient all-word averaging circuit 9.
The average correlation coefficient extracted from all recognition target words is calculated for each phoneme. This value is stored in the average correlation coefficient storage section 10. Next, if the word that is currently being uttered is one that will later be used for learning as part of the words to be recognized (a learning word), switch 7.
is connected to the terminal 8b, and the average correlation coefficient of each phoneme section is sent to the autocorrelation coefficient partial word averaging circuit 13, which calculates the average correlation coefficient extracted from the learning words for each phoneme. be done. This value is stored in the average correlation coefficient storage section 14. When all learning samples are uttered in a predetermined order and finished, the contents of the average correlation coefficient storage unit 10 are as follows.
The log cross section ratio calculation circuit 11 converts the value into a log cross section ratio value, which is stored in the all word learning log cross section ratio storage section 12 .

Similarly, the contents of the average correlation coefficient storage section 14 are converted into a log cross section ratio value by the log cross section ratio calculation circuit 15 and stored in the partial word learning log cross section ratio storage section 16 . The above operations are repeated for a predetermined number of speakers, for example, about 30 speakers, and the average correlation coefficient storage unit 10, the whole word learning log cross-sectional area ratio storage unit 12, and the partial word learning logarithmic cross-sectional area ratio The storage unit 16 separately stores values related to each speaker.

When the utterances of all predetermined speakers are completed, the value of the average correlation coefficient of each phoneme of each speaker stored in the average correlation coefficient storage section 10 is sent to the autocorrelation coefficient averaging circuit 17. For each phoneme, the average correlation coefficient is calculated for all speakers. This value is converted into a log cross-sectional area ratio value by the log cross-sectional area ratio calculation circuit 18,
It is stored in the average log cross-sectional area ratio storage section 19. Next, in order to obtain a conversion formula that estimates the all-word learning log-area ratio of all phonemes by linearly combining all the log-cross-sectional area ratios of all phonemes obtained by partial word learning, we calculate the all-word learning log-area ratio. The contents of the cross section ratio storage section 12 and the contents of the partial word learning log cross section ratio storage section 15 are sent to the normal matrix calculation circuit 22, and are calculated for all combinations between phonemes stored in both storage sections. A normal matrix calculation is performed. The S-th component of phoneme 1 (1
KuSkuτ. ．． AO), and the elements φ,,s(M.
．． n), (a) KumkuτMaxlOkunkuτ. ．． Ax+
1) is defined by the following equation. However, Zjl(m) indicates the m-th component of the log cross-sectional area ratio of phoneme j by speaker 1, and Z, , (0)=1. ) This normal matrix divides the S-th component Zll(S) of phoneme 1 (7) into the component Z of phoneme j stored in the partial word learning log cross-sectional area ratio storage unit 16.
Bn. It is a coefficient matrix of simultaneous linear equations that determine (S).

And B. n(S) is the element φI,s(
M. ．． It is determined by simultaneous linear equations with n) as a coefficient.

The values of the normal matrix are sent to the normal equation calculation circuit 23, and a solution calculation of the simultaneous linear equations using the coefficients φ, , s(M.n) is performed.

A sweep method or the like is used for this solution calculation. The resulting solution is here Bm(S), (0km
It is hail. The above calculations are performed for all combinations of degrees and phonemes from the 1st to γMaxth, and these results are stored in the conversion formula storage unit 24.

At the same time, the contribution rate of (Zil(m))(a)kmkuτMa,) to Zll(S) is calculated and stored in the weight storage unit 28.

The contribution rate Wij(S) is defined by the following formula. However, 71
(S) is the average value for all speakers of Zll(S).

Next, when learning the voice characteristics of a new speaker who was not used in the above processing, the switch 7 is connected to the terminal 8b, and the switch 20 is connected to the terminal 21b. Some learning words among the words to be recognized are uttered in the same predetermined order as in the above process.

This audio wave is input to the audio signal input terminal 1 in the same manner as described above, and is passed through the low-pass filter 2 and the analog-digital signal input terminal 1. The signal is converted into a time series of autocorrelation coefficients through a tal conversion circuit 3 and an autocorrelation coefficient calculation circuit 4. Furthermore, the phoneme segmentation circuit 5
Accordingly, the phoneme segment is segmented into each phoneme segment, and the intra-segment autocorrelation coefficient averaging circuit 6 calculates the average correlation coefficient of each phoneme segment. This value is the autocorrelation coefficient partial word averaging circuit 1.
3, and the average correlation coefficient across all training words is calculated for each phoneme. This value is stored in the average correlation coefficient storage unit 14, and when all learning words have been uttered, the log cross section ratio 4 calculation circuit 15 calculates
It is converted into a logarithmic cross section ratio value and stored in the partial word learning log cross section ratio storage section 16. This value is then sent to the standard pattern calculation circuit 25, and at the same time, the conversion formula storage section 24
, the contents of the average log cross section ratio storage section 19, and the contents of the weight storage section 28 are sent, and when all the words to be recognized are uttered for all components of the log cross section of each phoneme. A value corresponding to the value obtained in is calculated. The following formula is used for this calculation.
\A
−1. However, Zjl is the target log cross-sectional area ratio vector of phoneme 1 of speaker 1 (Zll=
(Zll(1), Zll(2), , Zil(γMa.
)"), Zil is the log cross section ratio vector of phoneme j of speaker 1 stored in the partial word learning logarithm 7 cross section ratio storage unit 16, ΩL is the set of phonemes included in the learning word, ψI
j is the 1st to τ, 30th component of the pairwise cross-sectional area ratio of phoneme 1 for the combination of phoneme 1 when all words are uttered and phoneme j when only the learning words are uttered. The transformation formula for is arranged in order and made into a matrix, r
is a weight, and a value such as 0.5 is empirically used.

If phoneme 1 or a phoneme similar to it is included in ΩL, Yil is determined by using the value of the log cross section ratio of the phoneme stored in the partial word learning log cross section ratio storage unit 16. If not, the average log cross section ratio storage unit 1
The value of the log cross-sectional area ratio of phoneme 1, which is averaged over a large number of speakers stored in 9, is used. The above equation (7) is 2
It consists of two terms.

The first term is a term that uses the transformation matrix ψIj to associate the log cross-sectional area ratio Z of various phonemes extracted from the learning words with the log cross-sectional area ratio Zll of each phoneme to be estimated.
Rather than simply estimating Zll as the average of the estimated values for each phoneme, the reliability of the estimated value is increased by performing a weighted average using the contribution rate Wij (vector) corresponding to the reliability of the estimation. . In order to avoid extreme fluctuations in the estimated value due to the estimation error in the first term, the second term is the log cross-sectional area ratio of the phoneme directly extracted from the learning word or the average value of many speakers, and the first term There is a method to stabilize the estimated value by calculating a weighted average with the estimated value of . The value of the standard pattern Zjl of phoneme 1 of the speaker 1 calculated in this way is calculated by the maximum likelihood spectral parameter calculation circuit 2.
6 and is converted into the value of the maximum likelihood spectral parameter.

The value of the maximum likelihood spectral parameter can be calculated by converting the log cross section ratio into a PARCOR coefficient, converting this into a linear prediction coefficient, and finding the sum of the products. Details of the method are described in literature (1) and others. References (1) Itakura: Communication based on voice characteristic parameters, 1971 Joint Conference of the Four Electrical Engineers of Japan,
The maximum likelihood spectral parameters of each 219 phoneme are stored in the standard pattern storage unit 27 as a standard pattern for word speech recognition.

With this structure, all the words to be recognized are uttered based on the log cross-sectional area ratio of the phonemes extracted from a small number of learning words, using a standard pattern with phonemes as units. Since the value can be adapted to the speaker as a value close to the value obtained at the time, it becomes possible to perform machine recognition of speech extremely efficiently and with a high recognition rate. As explained above, according to the present invention, the spectrum (log cross-sectional area ratio) of each phoneme when a small number of learning words are uttered, and the average of each phoneme when all words to be recognized are uttered. A general relationship between the spectrum (log cross section ratio) is prepared as a conversion formula, and the value obtained by applying this conversion formula to the spectrum of each phoneme obtained when a small number of learning words are uttered is calculated as the value of each phoneme. Because it is used as a standard pattern, it is possible to obtain a high recognition rate using a small number of learning words even in speech recognition that targets a large number of words.
Since it is possible to provide a speech signal recognition device that is extremely easy to use and inexpensive, the present invention can provide a method for learning individual differences in speech recognition that is extremely useful.

[Brief explanation of the drawing]

The figure is a block diagram of an embodiment of the invention. DESCRIPTION OF SYMBOLS 1...Audio input terminal, 2...Low pass filter, 3...Analog-digital converter, 4...Autocorrelation coefficient calculation circuit, 5... ...Phonological section segmentation circuit, 6...Intra-segment autocorrelation coefficient averaging circuit, 7...Switch, 8a, 8b...
Switch terminal, 9...Autocorrelation coefficient all word averaging circuit, 10...Average correlation coefficient storage unit, 11.
... Log cross section ratio calculation circuit, 12 ... All word learning log cross section ratio storage unit, 13 ... Auto correlation coefficient partial word averaging circuit, 14 ... ... Average correlation coefficient storage section, 15 ... Logarithmic cross section ratio calculator/calculation circuit,
16... Partial word learning log cross section ratio storage unit, 1
7...Autocorrelation coefficient averaging circuit, 18...Log cross section ratio calculation circuit, 19...Average log cross section ratio storage unit, 20...Switch, 21a, 21b...
...Switch terminal, 22...Normal matrix calculation circuit,
23... Normal equation calculation circuit, 24...
・Conversion formula storage unit, 25...Standard pattern calculation circuit, 26...Maximum likelihood spectral parameter calculation circuit, 2
7...Standard pattern storage section, 28...
Weight storage section.

Claims (1)

[Claims] 1. Means for calculating autocorrelation coefficients up to a finite degree from speech waves, means for segmenting speech waves into various phoneme intervals, means for averaging the autocorrelation coefficients within each segment, and storage for storing them. a means for calculating an average value of autocorrelation coefficients for all words to be recognized for each phoneme;
a storage unit that stores them; a means for calculating the average value of autocorrelation coefficients for each phoneme included in some learning words of the recognition target vocabulary; a storage unit that stores them; Means for converting correlation coefficients into log cross section ratios, and log cross section ratios converted from correlation coefficients averaged for all recognition target words and averaged for some learning words to be recognized. Means for calculating the numerical relationship between the correlation coefficient and the log cross-sectional area ratio as a multidimensional conversion formula for all phoneme combinations, a storage unit for storing this conversion formula, and recognition using this conversion formula. Based on the value of the log cross section ratio converted from the correlation coefficient averaged for some of the target learning words, the log cross section converted from the correlation coefficient averaged for all recognition target words. When a new speaker, who was not used to calculate the above conversion formula, utters some of the learning words to be recognized, from this speech wave Calculate the average correlation coefficient for these words with various phonemes, convert this value to a log cross-section ratio, and use this value and the conversion formula above to calculate the phonology included in the learning word and the phonology included. By calculating the value equivalent to the log cross-sectional area ratio converted from the correlation coefficient averaged for all recognition target words, the standard pattern with phonology as a unit can be adjusted to individual differences. A learning processing method for individual differences in speech recognition characterized by adaptive creation.