JPH0426480B2 - - Google Patents

Description

[Detailed Description of the Invention] "Industrial Application Field" This invention calculates the degree of similarity between the time series of features of input speech and the time series of the corresponding standard pattern speech.
The present invention relates to a speech recognition device that performs voice recognition based on calculating distances between these time-series elements.

"Prior art" In this type of conventional speech recognition device, the time series of input speech features 〓=〓 ₁ , 〓 ₂ , ...〓 _l and the time series of pattern speech features 〓 = 〓 ₁ , 〓 ₂ ,...〓 _n , the distance measure between the elements 〓 _i and 〓 _j was found using a spectrum in which the total power in each element was normalized to a constant value. For this reason, there was a drawback that the distance could become quite small even between elements with extremely different vocal powers, such as element _i being a vowel and element _j being a consonant.

In speech recognition, it is desirable to be able to recognize as many people's voices as possible using one standard pattern, but many conventional methods use distances based on spectra that are easily affected by individual differences, so vowels and consonants Errors such as erroneous recognition of the standard pattern were likely to occur, and the recognition rate significantly decreased, especially when the input speech was not from the same speaker as the standard pattern.

To address this problem, attempts have been made to use a matching distance measure between elements of speech time series that combines spectral information and power information (for example, Ref. Aikawa, Kano, Furui: Distance weighted by power information). (See Acoustical Society of Japan Speech Study Group Materials, S81-59, 1981). However, in this method, it is necessary to deal with fluctuations in power information due to the input level of the voice, so it is necessary to check the maximum and minimum values of the power of the entire voice section and normalize the voice power so that these are constant values. It was hot.
For this reason, it is necessary to check the changes in voice power until the end of the voice section or over several seconds, and then start calculating the distance measure, which may cause a time delay before recognition results are obtained, or The disadvantage was that a high-speed element had to be used to calculate the scale.

SUMMARY OF THE INVENTION An object of the present invention is to provide a word speech recognition device that is capable of recognizing speech of unspecified speakers more accurately and without delay than conventional methods.

"Means for Solving the Problem" According to the present invention, the distance of the linear regression coefficient of the temporal waveform of audio power is used in conjunction with the spectral distance, which has little individual difference and is not affected by fluctuations in the audio input level. . In other words, the inter-element spectral distance D _s of the feature time series of the input speech and the standard pattern speech is calculated, and the time waveform of the linear regression coefficient is calculated from the time waveform of the audio power for the input speech and the standard pattern speech at all points in time. and use these linear regression coefficients to calculate the power regression coefficient distance D _p between the input voice and the standard pattern.
seek. Using the inter-element matching distance determined from the spectral distance D _s and the power regression coefficient distance D _p , the degree of similarity between the input speech and the standard pattern speech is calculated by time-normalized matching.

"Embodiment" FIG. 1 shows an embodiment of the present invention, in which the input audio signal applied to the audio input terminal 1 is first subjected to a speech section detection circuit 2, in which silent (noise) sections are removed and an actual speech section is obtained. only are extracted. To detect this voice section, several well-known methods can be used, such as the short-time power of the input voice signal wave, the time period during which the power of a certain value or more continues. The signal wave of the detected voice section is sent to the linear prediction analysis circuit 3, where it is converted into a linear prediction coefficient and a time waveform of power. Since this technique is already known (for example, literature, Itakura, Saito: Estimation of speech spectral density and formant frequency using statistical methods, Journal of the Institute of Electronics and Communication Engineers, 53-A, 1,
(Refer to p. 35, 1970), details are omitted, but basically, first pass it through a low-pass filter, then sample and quantize it, multiply the waveform in a short period by a Hamming window, etc., cut it out, and sum the products. The power and correlation coefficient are calculated by the calculation. For example, a value such as 30 ms is used as the length of the Hamming window, and a value such as 10 ms is used as the period for updating this window. A linear prediction coefficient is extracted from this correlation coefficient by solving an algebraic equation through repeated calculation processing. The linear prediction coefficients are calculated, for example, from the 1st order to the 10th order.

The time waveform of the extracted linear prediction coefficient is converted into a linear prediction cepstrum coefficient by the cepstrum conversion circuit 4. Since this technique is already publicly known (for example, see literature, Saito, Nakata: Fundamentals of Speech Information Processing, Ohmsha, Chapter 7, P102, 1981), the details will be omitted, but the calculation of the recursive formula using linear prediction coefficients will be omitted. As a result, linear predictive cepstral coefficients (hereinafter simply referred to as cepstral coefficients for simplicity) can be easily obtained. The extracted cepstrum coefficients are temporarily stored in the feature parameter register 5.

On the other hand, the time waveform of the power, which is the other characteristic extracted by the linear prediction analysis circuit 3, is obtained by logarithmically converting the waveform of a certain time length section every extraction period (10ms in the above example), and then register 6
The contents of this register 6 are sent to a regression coefficient calculation circuit 7, where a linear regression coefficient is calculated. The length of the time waveform input to this register 6 and regression coefficient calculation circuit 7 is, for example,
Use a value like 50ms. When expressed as a logarithmic power time waveform x _j (j=-M, . . . M), this linear regression coefficient a (hereinafter referred to as a power regression coefficient) can be obtained by the following calculation.

a=( _M Σ ^j=-M x _j・j)/( _M Σ ^j=-M j ² )...(1) The power regression coefficient is input to the regression coefficient calculation circuit 7 which is updated every cycle as described above. It is calculated accordingly and stored in the feature parameter register 5 together with the cepstral coefficients.

The switch 8 is a switch for selecting a learning mode or a recognition mode. First, the switch 8 is connected to the terminal 8a, and then the user who inputs the voice to be recognized later, or multiple users different from the user, A feature parameter waveform consisting of a cepstrum coefficient and a power regression coefficient is obtained from the speech for each recognition target vocabulary, stored in a feature parameter register 5, and then input to a standard pattern storage section 9, where it is stored as a standard pattern for that vocabulary.

Switch 8 for subsequent voices to be recognized.
is connected to the terminal 8b, and the contents of the feature parameter register 5 are transferred to the time normalization matching circuit 10.
Enter. At the same time, standard patterns corresponding to each vocabulary are read out one by one from the standard pattern storage section 9 and input to the time normalization matching circuit 10. The time normalization matching circuit 10 calculates the degree of similarity of feature parameters between the standard pattern and the input speech.

The rate of speech production changes both partially and completely each time the same speaker utters the same word repeatedly, so to compare the two, we need to use common sounds (phonemes).
It is necessary to appropriately expand or contract one time axis non-linearly to match the other time axis so that the two time axes correspond to each other, and then compare the feature parameters at corresponding points in time.
Optimization using dynamic programming can be used as a technique for non-linearly expanding or contracting the time axis of one side so that it best matches the other (so that the degree of similarity between the two is greatest). (Reference: Sakoe, Chiba: Continuous word recognition based on temporal normalization of speech using dynamic programming, Journal of the Acoustical Society of Japan, 27, 9, P483, 1971).

Also in the apparatus of the present invention, the time normalization matching circuit 10 performs dynamic programming calculations. The cepstrum coefficient at a certain point k in the standard pattern is C ^R _ki (1ip, p is a value such as 10 as mentioned above), the power regression coefficient is a ^R _k , and the cepstrum coefficient at a certain point l in the input audio is
C ^I _li (1ip), and the power regression coefficient is expressed as a ^I _l . Here, the standard pattern at time k and time l regarding the cepstral coefficient and the power regression coefficient, respectively.
distance from the input voice (a numerical value indicating that the smaller the degree of similarity is) D _s (k, l), D _p (k, l)
The following values are used as

D _s (k, l)= _p Σ ⁱ⁼¹ (C ^R _ki − C ^I _li ) ² …(2) D _p (k, l)=(a ^R _k −a ^I _l ) ² …(3 ) Next, D(k, l) is obtained by weighted averaging of both of them as follows, and this value is used as the matching distance between elements of the standard pattern at time point and the input voice at time l, and the dynamic programming method is used. Perform calculations.

D(k,l)=√ _s (,)+(1−) _p (,)
...(4) The weight W used in this formula has a value of 0 to 1, and this value is set to an appropriate value based on the results of preliminary experiments to obtain a relatively high recognition accuracy and is stored in the weight register. Save it for 11.

The inter-element matching distance between the standard pattern and the input voice at corresponding points when the time axes are matched so that the degree of matching between the standard pattern and the input voice is the best through dynamic programming calculations is calculated over the entire voice interval. Calculate the average value for. This value will be referred to as the overall distance between the standard pattern and the input voice.
The total distance between the standard pattern corresponding to each vocabulary and the input speech is input to the comparison circuit 12, and the logic circuit determines the vocabulary with the smallest total distance among all these total distances. This determination result is output from the output terminal 13.

Incidentally, the temporal waveform of speech power has a basic property that it is high in the vowel part and low in the consonant part, and this property does not change even if the speaker is different. Figure 2 shows the logarithmic power time waveform of the word "Sapporo" uttered twice by each of the four speakers.The logarithmic power time waveform is normalized so that the maximum and minimum values are constant. It shows. This second
As can be understood from the figure, the power time waveform does not differ much even when the speaker changes, and it changes relatively smoothly over time. If we find the linear regression coefficient of the waveform, that is, the slope when linearly approximated, this value is common to different speakers because it is unaffected by the principle of linear regression coefficient even if the overall temporal waveform increases or decreases by a certain amount. It is possible to extract stable word features that are not affected by fluctuations in utterance level. Therefore, if time-normalized matching of the standard pattern and the input voice is performed by combining the power regression coefficient with the cepstrum coefficient as in this example, parts that are similar in both spectrum (cepstrum) and power will be matched, and vowel It is possible to avoid matching with consonants and improve the recognition rate. As a result of this structure, it is possible to use stable features included in the power time waveform without normalizing the power time waveform by examining the maximum and minimum power values in the entire speech interval. As a result, it is possible to realize a word speech recognition device that can start calculations for recognition immediately when speech is input, and can output recognition results with high precision for anyone's voice without any time delay. According to previous experiments, when 100 words of city names were recognized and the standard pattern was the voice of one speaker different from the person, the recognition rate with a conventional device using only cepstral coefficients was 85.5%. In contrast, the device of this example achieved a recognition rate of 89.3%, confirming that the present invention is superior.

By calculating the linear regression coefficient b (referred to as the cepstrum regression coefficient) of the cepstrum coefficients, and performing time-normalized matching of the similarity between the input speech and the standard pattern speech using the cepstral coefficients, cepstral regression coefficients, and power regression coefficients. , an even higher recognition rate can be obtained.

FIG. 3 shows this example, and parts corresponding to those in FIG. 1 are given the same reference numerals. The cepstrum coefficient C _o calculated by the cepstrum conversion circuit 4 is directly supplied to the feature parameter register 5, and the time waveform of this cepstrum coefficient C _o is temporarily stored in the cepstrum register 14 in sections of a certain time length at regular intervals. The contents of this register 14 are sent to a regression coefficient calculating circuit 15, where a linear regression coefficient (cepstrum regression coefficient) is calculated. This cepstrum register 14 and regression coefficient calculation circuit 1
For example, the length of the time waveform input to 5 is
50ms, the update cycle is, for example,
Use a value like 10ms. If the time waveform of the cepstrum coefficient is expressed as y _j (j=-M,...M),
This cepstral regression coefficient b can be obtained by the following calculation.

b=( _M Σ ^j=M y _j・j)/( _M Σ ^j=M j ² ) ...(5) The cepstrum regression coefficient b is a regression that is updated every 10ms for the cepstrum coefficient of each order. The cepstrum regression coefficient b is calculated according to the input of the coefficient calculation circuit 15, and is sent to the feature parameter register 7 and stored together with the cepstrum coefficient as a 2p-dimensional feature parameter. In the time normalization matching circuit 10, if the cepstrum coefficient and cepstrum regression coefficient at a certain point k of the standard pattern are expressed as r _ki (1i2p), and the cepstrum coefficient and cepstrum regression coefficient at a certain point l of the input audio are expressed as x _li (1i2p), Here, the following value is used as the distance between the two (a numerical value indicating that the smaller the degree of similarity is).

d=1/2p _2p Σ ⁱ⁼¹ w _i ² (r _ki −x _li ) ² ...(6) The reason for i=2p is that the order of the cepstral coefficient is P, and the order of the cepstral regression coefficient is P. , this is because the total order of both is 2P. Here, w _i is a numerical value indicating a predetermined weight for each coefficient, and this value is set to an appropriate value based on the results of preliminary experiments to obtain relatively high recognition accuracy. Store it in. As shown in equation (6), the distance d is calculated as the sum of squares of the difference between the input speech and the standard pattern for the P-th order cepstrum coefficient and the P-th order cepstrum regression coefficient at the same time, that is, the cepstrum coefficient and cepstral regression coefficients, which have different properties, are used together, and in order to balance them, w _i is weighted, so w _i
At least two values are used: w _a used when calculating the cepstrum coefficient, and w _b used when calculating the cepstrum regression coefficient. These weights w _a to w _b are weight register 1
Save it to 6.

In the time normalization matching circuit 10, the distance d(k,l) between the standard pattern at time point k obtained by equation (6) and the input voice at time point l is further calculated as D _s (k, l) in equation (4).
This equation (4) is calculated using Other operations are the same as in the case of FIG.

Note that although cepstral coefficients are used as voice feature quantities, linear prediction coefficients, formant frequencies, Percoll coefficients, logarithmic cross-sectional area ratios, number of zero crossings, etc. may also be used.

"Effects of the Invention" As explained above, according to the present invention, input speech and standard pattern speech are matched using distances consisting of power regression coefficients and spectral distances, so spectral distances alone may cause recognition errors. This method has the advantage that recognition ability can be improved in easy speaker-independent word speech recognition, and there is no need for normalization calculation of the absolute value of power, so there is no time delay in recognition calculation.

[Brief explanation of drawings]

FIG. 1 is a block diagram functionally showing an embodiment of the word speech recognition device of the present invention, FIG. 2 is a diagram showing the time pattern of the speech logarithmic power of the word "Sapporo", and FIG. FIG. 2 is a block diagram functionally showing an embodiment. 1: Audio input terminal, 2: Audio section detection circuit,
3: Linear prediction analysis circuit, 4: Cepstrum conversion circuit, 5: Feature parameter register, 6: Power register, 7: Regression coefficient calculation circuit, 8: Switch,
9: standard pattern storage section, 10: time normalization matching circuit, 11: weight register, 12: comparison circuit, 13: output terminal.

Claims (1)

[Claims] 1. Means for determining the spectral distance D _s between the elements of the feature time series of the input voice and the standard pattern voice, respectively, from the time waveform of the voice power for the input voice and the standard pattern voice, respectively. means for deriving a time waveform of the linear regression coefficient for all time points; means for using the linear regression coefficient to determine a power regression coefficient distance D _p between elements of the input speech and the standard pattern; and the spectral distance D _s and the A word speech recognition device comprising means for calculating the degree of similarity between the input speech and a standard pattern speech by time normalized matching using an inter-element matching distance determined from a power regression coefficient distance D _p .