JP3298658B2 - Voice recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition system, and more particularly to a voice recognition system which calculates a similarity or a distance by matching each template of a dictionary registered in advance with an input pattern. The present invention relates to a noise-tolerant speech recognition method for improving a recognition rate under high-level noise by adding a noise pattern at the time of recognition to each template on the dictionary side in a speech recognition method for recognizing.

[0002]

2. Description of the Related Art As a noise-tolerant speech recognition method conventionally proposed, a noise component of a noise section before and after a speech section is extracted from a speech signal at the time of recognition in which noise is superimposed. The spectrum subtraction method of subtracting the noise component from the voice signal on which noise is superimposed, or the noise power spectrum at the time of recognition extracted as described above is added to each template on the dictionary side registered in advance. There is a template addition method to be added, and its effectiveness has been conventionally confirmed.

FIG. 9 is a diagram showing a speech recognition system for performing noise processing by the above-described method.
1 is a voice input unit, 202 is a first frequency analysis unit, 20
2 ′ is a second frequency analysis unit, 204 is a first data compression unit, 204 ′ is a second data compression unit, 205 is a dictionary,
06 is a noise adding unit, 207 is a matching unit, and 208 is a nonlinear processing unit.

[0004] In FIG. 1, at the time of voice registration, when a voice signal is input from a voice input unit 201, a voice is analyzed at a first frequency analysis unit 202 at each frequency or each frequency band (hereinafter referred to as a channel). The magnitude of the power spectrum of the sound is output for each channel.
An output of the first frequency analysis unit 202 is logarithmically converted by calculating a logarithm in a first data compression unit,
It is stored in the dictionary 205 as a template.

At the time of voice recognition, the voice input unit 20
When an audio signal is input from No. 1, similarly to the above, the first frequency analysis unit 202 analyzes the audio for each channel and outputs the magnitude of the power spectrum of the audio for each channel. The output of the first frequency analysis unit 202 is logarithmically converted by calculating the logarithm in the first data compression unit 204, and is provided to the matching unit 207.

On the other hand, regarding the noise section, the voice input unit 2
01, the second frequency analysis unit 20
At 2 ', the noise is analyzed for each channel,
The signal is provided to the noise adding unit 206. Further, as described above, since each template stored in the dictionary 205 is logarithmically stored at the time of registration, it is stored in the dictionary 205.
At 8, the sound power is returned to the dimension of the voice power using an exponential function that is an inverse logarithmic function, and is provided to the noise adding unit 206.

[0007] The noise adding unit 206 adds a noise component to the template stored in the dictionary by adding the output of the nonlinear processing unit 208 and the output of the second frequency analysis unit 202 '. The output of the noise adding unit 206 is logarithmically calculated in a second data compression unit 204 ′, and is supplied to a matching unit 207. The matching unit 207 performs matching between the input voice pattern output from the first data compression unit 204 and the template on the noise-processed dictionary side output from the second data compression unit 204 ′, and obtains a recognition result. Is output.

[0008]

In the above-described method, the input speech is subjected to noise processing only once, so that the processing is simple and is often used. On the other hand, only the noise component is used. However, there is a drawback that audio components are also drawn. In addition, in the above-mentioned method, since noise components are added to the dictionary side instead of subtracting noise components, there is no deformation of the spectrum pattern which is not essential, but the noise processing needs to be performed by the number of dictionary templates. And, despite its effectiveness, was not practical.

[0009] That is, in the example shown in FIG. 9, the nonlinear processing in the nonlinear processing unit 208 is performed once in order to return each template on the dictionary side to the dimension of the audio power, and the data compression unit 204 performs the comparison in the non-linear processing unit 208. The nonlinear processing needs to be performed once, and the nonlinear processing needs to be performed twice for each template. Therefore, assuming that the dictionary has N templates, 2N nonlinear processing and N additions are required, and the processing amount is large.
There is a problem that real-time processing is difficult.

Further, in the conventional speech recognition apparatus, the output of each frequency-analyzed channel is logarithmically converted for the following reason. 1) It considers that it corresponds to Weber-Fechna's law that the intensity of the sense is almost proportional to the intensity of the sound pressure. 2) At the time of matching, low and high power levels can be calculated with the same weight. 3) Data compression can be performed without reducing the dynamic range. 4) Spectral conversion processing that must be performed by multiplication and division, such as power normalization, can be performed only by addition and subtraction, and the calculation speed can be increased.

However, regarding the above 1) to 3),
Any function that has the same characteristics as logarithm may be used.
Regarding 4), if it is not necessary to perform multiplication / division such as power normalization, the logarithm does not need to be obtained. Conversely, when the converted value of the sum / difference of the two power spectra is required. Logarithm has no advantage. The present invention has been made in view of the above-described drawbacks of the related art, and the noise is reduced by adding the extracted noise power spectrum at the time of recognition to each template on the dictionary side registered in advance. In a speech recognition method for recognizing superimposed speech, the amount of calculation is significantly reduced by performing conversion using a non-linear function f (x) = b.exp (ax) + c instead of logarithmic conversion. It is possible,
It is another object of the present invention to provide a speech recognition system capable of satisfactorily recognizing speech even under high-level noise.

[0012]

FIG. 1 is a diagram illustrating the principle of the present invention. The present invention is configured as shown in FIG. 1 in order to solve the above-mentioned problem, and the invention of claim 1 of the present invention provides a voice input unit 1 for converting a generated voice acoustic signal into an electric signal.
First and second frequency analysis units 2 and 2 ′ for performing frequency analysis of an audio input signal and outputting an input audio pattern including analysis data of a plurality of channels for each analysis frame;
First and second data compression units 4 and 4 'for compressing the frequency pattern analyzed by the first and second frequency analysis units 2 and 2' while maintaining a dynamic range by non-linear conversion; A dictionary 5 for storing the template created from the data after compression, and a noise addition unit 6 for adding a noise component obtained from the output of the second data compression unit 4 ′ to the template in the dimension of the power spectrum during speech recognition. And a collating unit 7 for collating the input speech pattern compressed by the first data compression unit 4 with the output of the noise adding unit 6 and calculating the similarity or distance between the two. In the above, the following equation is used as a nonlinear function in the first and second data compression units 4 and 4 ′, and f (x) = b · exp (ax) + c (where a, b and c are constants) ) The constant a nonlinear function, b, a constant determining unit 3 for determining a c provided, using the above non-linear function, a frequency pattern in which the first and second frequency analysis unit 2, 2 'outputs the first
And data compression by the second data compression units 4 and 4 '.

According to a second aspect of the present invention, in the first aspect of the invention, the noise adding section 6 adds noise to the template read from the dictionary 5 by the following equation. x3 = (x1-c) · (x2-c) / b + c where: x1: the value of the template read from the dictionary 5 x2: the noise component obtained from the data compression unit 4 ′ x3: the output of the noise addition unit 6 According to a third aspect of the present invention, in the first or second aspect, the maximum error between the nonlinear function f (x) and the logarithmic function log (x) within the range of x handled by the nonlinear function f (x). Are determined so as to determine constants a, b, and c that minimize.

According to a fourth aspect of the present invention, there is provided the first or second aspect.
In the invention of the above, within the range of x handled by the nonlinear function f (x), the nonlinear function f (x) and the logarithmic function log
(A), a constant that minimizes the integral value of the absolute value of the difference from
It is configured to determine b and c. According to a fifth aspect of the present invention, in the first or second aspect, within the range of x handled by the nonlinear function f (x), the square of the nonlinear function f (x) and the logarithmic function log (x) is set. The configuration is such that constants a, b, and c that minimize the integrated value of the error are determined.

[0015] The invention of claim 6 is claim 1 or claim 2.
In the invention of the above, within the range of x handled by the nonlinear function f (x), the nonlinear function f (x) and the logarithmic function log
(X) is subjected to Taylor expansion, and the constants a, b, and c are determined by simultaneous equations for the constants a, b, and c when the coefficients of each term up to the three terms are assumed to be equal. It is.

[0016] The invention of claim 7 is claim 1 or claim 2.
In the invention, the recognition result counting section is provided to count the recognition results in the matching section 7, and the constants a, b, and c of the nonlinear function are changed based on the recognition rate to determine the optimum constant. It is. The invention of claim 8 is claim 1,
2, 3, 4, 5, 6, or 7, wherein the constant c of the nonlinear function is set to zero.

According to a ninth aspect of the present invention, in the invention of the eighth aspect, the template read out from the dictionary 5 is multiplied by a noise component obtained from the second data compression unit 4 ′, so that the noise addition unit 6 It is configured to obtain an output. According to a tenth aspect of the present invention, in storing the template in the dictionary 5, the constant b of the non-linear function
, The template value is divided.

According to an eleventh aspect of the present invention, in the ninth aspect of the invention, the input voice pattern compressed by the first data compression unit 4 is multiplied by a constant b of a non-linear function, and the result is supplied to the matching unit 7. It is what was constituted. Claim 12
According to the ninth aspect of the present invention, the noise pattern compressed in the second data compression section 4 'is divided by a constant b of a nonlinear function, and the result is given to the noise addition section 6. is there.

[0019]

According to the first and second aspects of the present invention, at the time of recognizing the voice, the input voice is frequency-analyzed by the frequency analysis unit 2, and the nonlinear function f
Data compression is performed according to (x) = b · exp (ax) + c, and the result is provided to the matching unit 7.

On the other hand, the noise signal is frequency-analyzed by a frequency analysis unit 2 ′, data is compressed by a non-linear function f (x) in a data compression unit 4 ′, and given to a noise addition unit 6. The noise adding unit 6 adds a compressed noise component to the template read from the dictionary 5 and provides the template to the matching unit 7.

The collating unit 7 calculates a similarity or a distance between the output of the noise adding unit 6 and the output of the data compressing unit 4, and outputs a recognition result. Nonlinear function f as a conversion function
Since (x) = b.exp (ax) + c is used, the noise component of the power spectrum dimension can be added only by the calculation of the four arithmetic operations, and the amount of calculation can be greatly reduced.

Further, by determining the constants a, b, and c of the nonlinear function according to the method of claims 3 to 6, within the range of x handled by the nonlinear function f (x), the nonlinear function f ( x) = b.exp (ax) + c can be approximated to a logarithmic function, and the same characteristics as when a logarithmic function is used can be obtained. Further, a recognition result counting unit is provided to count the recognition results in the collating unit 7, and based on the recognition rate, constants a and
By changing b and c, it is possible to perform data compression with the best recognition rate in the matching unit 7.

Further, by setting the constant c of the non-linear function to zero, the addition of the constant c in the data compression units 4 and 4 'and the addition of a noise 6, the subtraction of the constant c is not required, and the calculation amount is further reduced. As in the invention of claim 9, the noise component obtained from the second data compression unit 4 'is added to the template read from the dictionary 5. By multiplying, the output of the noise adding unit 6 can be obtained.

Even when the constant c is set to zero, when the distance between the two patterns is calculated by the collating unit 7, the calculated city distance does not affect the value of the calculated distance. . The difference in the magnitude of the absolute value due to the constant c is not essential even in the Euclidean distance for calculating the square difference between two patterns. Further, according to the tenth aspect of the present invention, when storing the template in the dictionary 5, the template value is divided by the non-linear function constant b, so that the output of the data Is required to be divided by the constant b, but only the multiplication needs to be performed in the noise adding unit 6 at the time of speech recognition. Since the division is unnecessary, the calculation amount at the time of speech recognition can be further reduced.

Further, as in the eleventh aspect of the present invention, the first
By multiplying the input voice pattern compressed in the data compression unit 4 by a constant b of a non-linear function and providing the result to the collation unit 7, the output of the data compression unit 4 during voice recognition is a constant. b needs to be multiplied, but the noise adding unit 6 only needs to perform the multiplication and does not need to divide, so that one recognition can save the number of divisions corresponding to the number of templates. Can be further reduced.

Still further, as in the invention of claim 12,
By dividing the noise pattern compressed by the second data compression unit 4 ′ by a non-linear function constant b and providing the result to the noise addition unit 6, the data compression unit 4 Although it is necessary to divide the output by the constant b, it is sufficient to perform only the multiplication in the noise adding unit 6, and the division is unnecessary, so that the number of divisions corresponding to the number of templates can be saved for one recognition. The amount of calculation at the time of speech recognition can be further reduced.

[0027]

FIG. 2 is a diagram showing a first embodiment of the present invention. In FIG. 2, reference numeral 11 denotes an analog / digital converter,
12 is a voice section detector, and 2 and 2 'are first and second sections.
Frequency analysis section, 3 is a constant determination section, 4 and 4 'are first
And a second data compression unit, 5 is a dictionary, 61 is a noise addition unit, and 7 is a collation unit.

In FIG. 1, an analog / digital converter 11 is a means for converting an audio input signal into a digital signal, and its output is given to an audio section detector 12. The voice section detection unit 12 is a means for determining a voice section based on information such as the power of an input signal. The voice section detection unit 12 detects a voice section of the input voice signal, and the voice section is determined by the first frequency analysis unit 2. Given to.

In the recognition, the word spotting method (when comparing a pattern signal to be identified with a template, one signal is shifted with respect to the other signal, and the two signals are shifted at a position where the error is minimized. (A method of determining a pattern signal to be identified and a template distance by calculating the distance of the voice signal) is not necessarily required.

The first frequency analysis unit 2 and the second frequency analysis unit 2 ′ are means for performing frequency analysis on both the audio signal and the noise signal output from the audio section detection unit 12. A frequency analysis is performed using a 19-channel band-pass filter, and the power
Outputs the magnitude of the spectrum. First frequency analysis unit 2
And the output of the second frequency analysis unit 2 ′ is the data compression unit 4
And a non-linear function f
Data compression is performed by (x) = b.exp (ax) + c. The constants a, b, and c of the nonlinear function are determined by the constant determining unit 3.

The constant determining unit 3 calculates the nonlinear function f (x) = b ·
exp (ax) + c is the logarithmic function g conventionally used
This is means for determining constants a, b, and c so as to approximate (x) = logx, and the following method can be used as an approximation method. A method of determining constants a, b, and c so as to minimize the maximum error between f (x) and g (x). A method of determining constants a, b, and c that minimizes the integral value of the absolute value of the difference between f (x) and g (x). A method of determining constants a, b, and c that minimizes the integral value of the square error between f (x) and g (x). f (x) and g (x) are Taylor-expanded, and constants a, b, and c when coefficients of each term up to three terms are assumed to be equal
A method of determining three constants by a ternary simultaneous equation according to

FIG. 3 is a diagram showing a logarithmic function (a) and nonlinear functions (b) to (c) represented by the following equations. (A) f (x) = log (x) (b) f (x) = b · exp (ax) + c (when c = 0) (c) f (x) = b · exp (ax) −be ^a ((b)
In the case of c = −be ^a ) As is clear from the figure, by selecting the constants a, b, and c, it is possible to obtain a non-linear function approximating a logarithmic function.

The dictionary 5 shown in FIG. 2 stores a template input at the time of voice registration.
The template is read out. The noise adding unit 61 is a means for adding noise to the template read from the dictionary. The matching unit 7 calculates the distance from the voice pattern signal output from the data compression unit 4, and determines the recognition result of the input voice. Output means.

Next, the operation of the first embodiment shown in FIG. 2 will be described. When voice is registered, when the voice is input, the input voice is converted into a digital signal by the analog / digital converter 11 and the voice section is detected by the voice section detector 12. The output of the voice section detection unit 12 is analyzed by a frequency analysis unit 2 for each channel by a band pass filter, and the magnitude of the power spectrum is output.

The analysis result of the voice is given to the data compression unit 4, where the data is compressed by the nonlinear function f (x) = bexp (ax) + c, and the dictionary 5 is used as a template.
Registered in. At the time of voice recognition, when a voice is input, the input voice is converted into a digital signal by an analog / digital converter 11 and a voice section is detected by a voice section detector 12.

The speech signal on which noise is superimposed detected by the speech section detection unit 12 is subjected to frequency analysis by the frequency analysis unit 2 in the same manner as described above, and the data compression unit 4 performs a nonlinear function f (x) = b Data compression is performed by exp (ax) + c, and the result is provided to the matching unit 7. On the other hand, the noise signal is provided from the voice section detection unit 12 to the frequency analysis unit 2 ', and is subjected to frequency analysis in the same manner as described above, and is subjected to the nonlinear function f (x) = b.exp (ax) in the data compression unit 4'.
The data is compressed by + c, and is given to the noise adding unit 61.

The noise adding section 61 adds a compressed noise component to the template read from the dictionary 5 and supplies the template to the matching section 7. Here, s is the power spectrum of the voice, n is the power spectrum of the noise, and the converted value of the power spectrum of the voice by the nonlinear function f (x) = bexp (ax) + c is f ( s), assuming that the converted value of the noise power spectrum is f (n), the converted value f (s + n) of the power spectrum s + n in which the noise is superimposed on the voice can be obtained by the following equation.

F (n + s) = b · exp (an + as) + c = b · exp (an) · exp (as) + c = [{b · exp (an) + c} −c] · [｛b · exp (as ) + C｝ −c] / b + c = {f (n) −c} · {f (s) −c} / b + c Therefore, in the noise adding unit 61, the converted value f ( subtracting the constant c from s), subtracting the constant c from the noise conversion value f (n), which is the output of the data compression unit 4 ', dividing the product by the constant b, and further adding the constant c. Thus, a pattern signal on which noise is superimposed can be obtained, and this signal is provided to the matching unit 7.

The collating unit 7 compares a pattern signal obtained by adding noise at the time of recognition to the template output from the noise adding unit 61 and a voice pattern on which the analyzed noise output from the data compressing unit 4 is superimposed. For example, matching is performed using a DP (dynamic programming) matching method or the like, and a recognition result is output. In the present embodiment, a non-linear function f (x) = b · exp as a conversion function
Since (ax) + c is used, the noise component of the power spectrum dimension can be added only by the calculation of the four arithmetic operations.
It is not necessary to perform the non-linear processing twice per template as in the conventional example shown in (1), and the amount of calculation can be greatly reduced.

FIG. 4 is a diagram showing a second embodiment of the present invention. FIG. 4 shows a configuration in which a recognition result counting section 9 is added to the first embodiment shown in FIG. Is the same as the embodiment of FIG. In the figure, a recognition result counting unit 9 counts the matching result in the matching unit 7, and the nonlinear function f (x) = b · determined in the constant determining unit 3 so that the recognition rate in the matching unit becomes the best. This is a means for changing the constant of exp (ax) + c.

In the first embodiment shown in FIG. 2, the constant of the nonlinear function f (x) = b.exp (ax) + c is
Although the non-linear function f (x) is determined to approximate the logarithmic function g (x) = logx, the reason for using the logarithmic function as the non-linear function conventionally has no strict relation to the recognition rate. However, the logarithmic function is not always the best in terms of recognition rate.

Conversely, for the nonlinear function f (x), 3
Since the two constants can be changed, by changing these constants according to the recognition rate as in this embodiment, it is possible to perform data compression with the best recognition rate. FIG. 5 shows a third embodiment of the present invention. In this embodiment, the first and second data compression units 4 and 4 of the embodiment of FIG.
4 ′, the noise adding unit 61 is changed to first and second data compressing units 42 and 42 ′ and a noise adding unit 62,
Other configurations are the same as those of the first embodiment shown in FIG.

In the data compression units 42 and 42 'of this embodiment, the non-linear function f (x) = b.exp (ax) + c
In the noise adding unit 62, the addition of c is eliminated, and the subtraction of c is eliminated, so that the calculation amount is further reduced as compared with the embodiment of FIG. Even if c is omitted as in the present embodiment, the difference between the two patterns is calculated when the distance between the two patterns is calculated in the matching unit 7, and thus does not affect the value of the calculated distance.

For example, assuming that the city area distance is used as the distance scale in the collating unit 7, the distance between the two patterns is calculated as follows. | F (x1) −f (x2) | = | b · exp (ax1) + c− (b · exp (ax2) + c) | = | b · exp (ax1) − (b · exp (ax2)) | As is apparent from the formula, in the distance calculation, the distance between the two patterns does not depend on c.

FIG. 6 is a diagram showing a fourth embodiment of the present invention. In this embodiment, the noise adding section 62 of the third embodiment shown in FIG. The configuration is the same as that of the third embodiment shown in FIG. 5 except that a matching unit 101 is added. 6, at the time of voice registration, the output of the data compression unit 42 is divided by a constant b in the data matching unit 101 and stored in the dictionary 5 as a template.

Therefore, as described above, s is the power spectrum of the voice, and the nonlinear function f (x) = b · exp
The converted value of the power spectrum of the voice by (ax) is f
(S), the dictionary 5 contains f (s) / b = exp (a
The template represented by s) is stored. Then, at the time of speech recognition, a template represented by f (s) / b = exp (as) is read from the dictionary 5, and the noise adding unit 63 outputs f (n) = b · e
Since it is multiplied by xp (an), the output of the noise adding unit 63 is as shown in the following expression.

F (n) · f (s) / b = exp (as) · b · exp (an) = b · exp (as + an) = f (n + s) The output of the noise adding unit 63 is sent to the matching unit 7. The distance from the output of the data compression unit 42 is obtained as in the case of the third embodiment, and the recognition result is output. In the present embodiment, it is necessary to divide the output of the data compression unit 42 by a constant b at the time of voice registration.
In, only multiplication needs to be performed, and division is not necessary.
The amount of calculation at the time of speech recognition can be further reduced.

FIG. 7 is a diagram showing a fifth embodiment of the present invention. In this embodiment, the noise adding section 62 of the third embodiment shown in FIG. The configuration is the same as that of the third embodiment shown in FIG. 5 except that a matching unit 102 is added. In FIG. 7, at the time of speech recognition, a template corresponding to f (s) = b · exp (as) is read from the dictionary 5, and f (n) = b output from the data compression unit 42 in the noise addition unit 63.・ Exp
(An), and the output of the noise adding unit 63 is as shown in the following expression.

F (n) · f (s) = b · exp (as) · b · exp (an) = b ² • exp (as + an) = b · f (n + s) On the other hand, the output of the data compression unit 42 Since the data matching unit 102 multiplies by the constant b, the data matching unit 102
Is b · f (n + s). The output of the noise adding unit 63 is supplied to the matching unit 7, where the distance from the output of the data matching unit 102 is obtained and the recognition result is output, as in the third embodiment.

In the present embodiment, it is necessary to multiply the output of the data compression unit 42 by a constant b at the time of speech recognition. However, in the noise addition unit 63, only the multiplication needs to be performed. With respect to the number of times of recognition, division of the number of times corresponding to the number of templates can be saved, and the amount of calculation at the time of speech recognition can be further reduced. FIG. 8 is a diagram showing a sixth embodiment of the present invention. In this embodiment, the noise adding unit 62 of the third embodiment shown in FIG. The other configuration is the same as that of the third embodiment shown in FIG.

In FIG. 8, at the time of voice recognition, the output f (n) = b.exp (an) of the data compression unit 42 'is divided by the constant b in the data matching unit 103 to obtain f (n) / b
= Exp (an) is obtained and given to the noise adding unit 63. In the noise adding unit 63, f (s) = b · exp (as) read from the dictionary 5 and the data matching unit 10
3 is multiplied by f (n) / b = exp (an), and the output of the noise adding unit 63 is as shown in the following equation.

F (n) · f (s) / b = b · exp (as) · exp (an) = b · exp (as + an) = f (n + s) The output of the noise adding unit 63 is sent to the matching unit 7. As in the case of the third embodiment, the distance from the output of the data matching unit 102 is obtained, and the recognition result is output. In this embodiment, at the time of speech recognition, it is necessary to divide the output of the data compression unit 42 by a constant b. However, in the noise addition unit 63, only the multiplication needs to be performed. Therefore, the number of divisions corresponding to the number of templates can be saved, and the amount of calculation at the time of speech recognition can be further reduced.

[0053]

As is apparent from the above description,
In the present invention, the nonlinear function f (x) is used as the conversion function.
Since b = exp (ax) + c is used, the noise component of the power spectrum dimension can be added only by calculation of the four arithmetic operations, and the amount of calculation can be greatly reduced.

Also, a recognition result counting section is provided to count the recognition results in the matching section, and the constants a, b, and c of the nonlinear function are changed based on the recognition rate, whereby the recognition rate in the matching section is optimized. Data compression can be performed. Further, by setting the constant c of the nonlinear function to zero, the addition of the constant c in the data compression unit, and
In the noise adding unit, the subtraction of the constant c is not required, and the calculation amount can be further reduced.

[Brief description of the drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a diagram showing a first embodiment of the present invention.

FIG. 3 is a diagram showing a logarithmic function and a non-linear function.

FIG. 4 is a diagram showing a second embodiment of the present invention.

FIG. 5 is a diagram showing a third embodiment of the present invention.

FIG. 6 is a diagram showing a fourth embodiment of the present invention.

FIG. 7 is a diagram showing a fifth embodiment of the present invention.

FIG. 8 is a diagram showing a sixth embodiment of the present invention.

FIG. 9 is a diagram showing a conventional example.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Voice input part 2, 2 'Frequency analysis part 3 Constant determination part 4, 4', 42, 42 'Data compression part 5 Dictionary 6, 61, 62, 63 Noise addition part 7 Collation part 9 Recognition result counting part 11 Analog / Digital conversion section 12 Voice section detection section 101, 102, 103 Data matching section

──────────────────────────────────────────────────続 き Continuation of the front page (51) Int.Cl. ⁷ Identification code FI G10L 21/02 (56) References Sadahiro Furui, Digital Speech Processing, Japan, Tokai University Press, September 25, 1985, p. 99-100 (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 15/02 G10L 15/06 G10L 15/10 G10L 21/02 JICST file (JOIS)

Claims (12)

(57) [Claims] An audio input unit for converting a generated audio sound signal into an electric signal, a frequency analysis of the audio input signal, and an output of an input audio pattern composed of analysis data of a plurality of channels for each analysis frame. The first and second frequency analyzers (2, 2 ′) that perform the analysis and the frequency patterns analyzed by the first and second frequency analyzers (2, 2 ′) are dynamically converted by nonlinear conversion.
First and second data compression units (4, 4 ') for compressing data while maintaining a range, a dictionary (5) for storing a template created from learning data after data compression, and a second dictionary for speech recognition. A noise adding unit (6) for adding a noise component obtained from the output of the data compression unit (4 ') to the template in the dimension of the power spectrum; and an input voice pattern compressed in the first data compression unit (4). And a collation unit (7) for comparing the output of the noise addition unit (6) with the output of the noise addition unit (6), and calculating a similarity or distance between the two. The following equation is used as the nonlinear function in (4, 4 ′), and f (x) = b · exp (ax) + c (where a, b, and c are constants). The constants a, b, and c of the nonlinear function are determined. Constant determining unit
(3), and the first and second data compression units (4, 4 ′) use the nonlinear function to convert the frequency patterns output by the first and second frequency analysis units (2, 2 ′) by the first and second data compression units (4, 4 ′). A speech recognition method characterized by data compression. 2. A speech recognition system according to claim 1, wherein said noise adding section adds noise to the template read from the dictionary according to the following equation. x3 = (x1-c) · (x2-c) / b + c where: x1: value of template read from dictionary (5) x2: noise component obtained from data compression unit (4 ′) x3: noise addition Output of part (6) 3. Within the range of x handled by the nonlinear function f (x), the nonlinear function f (x) and the logarithmic function log
Constants a, b, c so that the maximum error with (x) is minimized
The speech recognition method according to claim 1 or 2, wherein: 4. Within the range of x handled by the nonlinear function f (x), the nonlinear function f (x) and the logarithmic function log
3. The speech recognition system according to claim 1, wherein the constants a, b, and c are determined so that the integral value of the absolute value of the difference from (x) is minimized. 5. Within the range of x handled by the nonlinear function f (x), the nonlinear function f (x) and the logarithmic function log
3. The speech recognition method according to claim 1, wherein the constants a, b, and c are determined so that an integral value of a square error with (x) is minimized. 6. Within the range of x handled by the nonlinear function f (x), the nonlinear function f (x) and the logarithmic function log
(X) is subjected to Taylor expansion, and constants a, b, and c are determined by simultaneous equations for constants a, b, and c when the coefficients of the respective terms up to three terms are assumed to be equal. The speech recognition system according to claim 1 or 2. 7. A recognition result counting unit for counting recognition results in a collating unit, and changing constants a, b, and c of a nonlinear function based on the recognition rate to determine an optimum constant. The speech recognition device according to claim 1 or 2, wherein 8. The speech recognition apparatus according to claim 1, wherein a constant c of the nonlinear function is set to zero. 9. An output of the noise adding unit (6) is obtained by multiplying a template read from the dictionary (5) by a noise component obtained from the second data compression unit (4 '). 9. The speech recognition method according to claim 8, wherein: 10. The speech recognition method according to claim 9, wherein, when storing the template in the dictionary (5), the template value is divided by a constant b of a nonlinear function. 11. The method according to claim 9, wherein the input voice pattern compressed in the first data compression section is multiplied by a constant b of a non-linear function, and the result is given to a collation section. Voice recognition system. 12. The noise pattern compressed by the second data compression unit (4 ') is divided by a non-linear function constant b, and the result is given to the noise addition unit (6). Nine voice recognition methods.