JPH05165492A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To improve the rate of recognition for the recognizing device by improving the S/N of an input voice by executing a preliminary processing due to a spectrum transaction to the input voice, and executing collation even to noise not completely removed while using a standard pattern from voice data having a character close to a noisy environment. CONSTITUTION:This device is equipped with a voice input part 102 for the voice as a recognizing object, noise processing part 106 to remove the noise by analyzing the frequency of an input voice signal and subtracting an estimated noise spectrum from an analyzed voice spectrum, analysis part 108 to calculate the feature vector of the input voice based on the voice signal obtained at the noise processing part, standard pattern selection part 110 to store the prepared standard pattern by using noise superimpose voice data corresponding to the noisy environment, to calculate the noise information of residual noise not completely removed by the noise processing part 106 and to select the optimum standard pattern out of the plural standard patterns based on the noise information, and collation part 112 to recognize the input voice by collating the standard pattern selected by the standard pattern selection part 12 with a feature vector.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device, and more particularly to a voice recognition device having high noise resistance for enabling recognition under noise.

[0002]

2. Description of the Related Art In order to put a voice recognition device into practical use, it is essential to have a noise resistance technique that correctly recognizes even a voice uttered in a noisy environment. Therefore, conventionally, a close-talking type microphone has been used exclusively to prevent ambient noise other than voice from being picked up as much as possible. However, the close-talking type microphone makes the voice input environment very limited, and there is an increasing need for a voice recognition device using a hands-free type microphone, which has a wider use. One of the needs is, for example, a vehicle-mounted voice recognition device. In order to realize a speech recognition system using a hands-free microphone, it is necessary to remove noise from speech with noise pre-processed or to perform recognition correctly even if noise is mixed. ..

[0003]

Conventionally, as means for removing noise, a means for removing noise using an adaptive filter, a spectral subtraction method for subtracting a noise spectrum estimated from noise mixed in speech from an input speech spectrum, etc. There is. However, in an actual environment,
Stationary and non-stationary noises from various sound sources are mixed in the speech, and it is very difficult to remove all these noise components from the input signal.

Further, as a method of improving noise resistance of the pattern matching recognition method, a method of performing matching using a parameter which is not easily influenced by noise, a method of using a distance measure, and noise being superposed on a standard pattern in advance. There is a noise superposition method. In the noise superimposition method, a standard pattern on which noise is superimposed must be prepared in advance, and if the characteristics of the noise to be superimposed on the standard pattern and the conditions such as S / N are different from those at the time of input, Since there is a problem that the effect of noise resistance is small, it is necessary to prepare a plurality of standard patterns according to the noise environment. However, there is a problem that the required memory amount becomes very large by increasing the types of standard patterns.

Another problem in putting the voice recognition apparatus into practical use is the problem of processing amount. In general, speech recognition requires a very large amount of processing. Therefore, in order to secure a sufficient amount of processing for noise countermeasures with limited resources, it is necessary to reduce the total processing for speech recognition. .. In the entire speech recognition, the matching process between the feature sequence of the input speech and the standard pattern has a particularly large weight. Therefore, reduction of the processing amount of the collation processing unit is one of the major problems for practical use of the voice recognition device.

It is an object of the present invention to provide a voice recognition device which is further improved in noise resistance and which can obtain a high recognition rate even in noise.

Another object of the present invention is to reduce the amount of memory required when a plurality of types of standard patterns are prepared by the noise superposition method and the amount of processing required for collation processing. It is to provide a practical voice recognition device.

[0008]

In order to achieve the above object, a voice recognition apparatus according to the present invention provides a voice input section for inputting a voice to be recognized and an input voice signal obtained by the voice input section. A frequency analysis is performed, and a noise processing unit that removes noise by subtracting the estimated noise spectrum from the frequency-analyzed speech spectrum, and a feature vector of the input speech based on the speech signal obtained by the noise processing unit. It stores a plurality of standard patterns created by using the analysis unit to be obtained and noise-superimposed voice data corresponding to a presumed noise environment, and removes it from the voice signal obtained from the noise processing unit by noise processing. A standard pattern selection unit that obtains noise information about the residual noise that cannot be cut off, and selects an optimum standard pattern from the plurality of standard patterns based on the noise information. Those having a recognizing matching unit the input voice against the standard pattern selected and the feature vector by the standard pattern selection unit.

Another voice recognition device according to the present invention is a voice input section for inputting a voice to be recognized and a frequency analysis of an input voice signal obtained by the voice input section to obtain a voice spectrum obtained by the frequency analysis. A noise processing unit that removes noise by subtracting the estimated noise spectrum,
An analysis unit that obtains a feature vector of the input voice based on the voice signal obtained by the noise processing unit, a plurality of types of codebooks created according to a presumed noise environment, and a vector using the codebook. A plurality of types of standard patterns as a code sequence obtained by quantization are stored, and information about residual noise that cannot be completely removed by noise processing from the voice signal obtained from the noise processing unit is analyzed. A selection unit that selects the type of the codebook and the type of the standard pattern by using the result, and the feature vector obtained by the analysis unit,
A vector quantizer that obtains a voice feature pattern by performing vector quantization using the selected type codebook, and the input by comparing the voice feature pattern with the selected type standard pattern. And a collating unit for recognizing voice.

[0010]

The present invention can be modified in various ways, and the operation of a typical structure will be described.

The input speech signal obtained by the speech input section is first subjected to frequency analysis by the noise processing section, noise is removed by subtracting the estimated noise spectrum from the frequency-analyzed speech spectrum, and the S / N ratio of the signal is reduced. Is improved. Next, the standard pattern selection unit obtains noise information about the residual noise that could not be removed by the noise processing unit, and created using noise-superimposed voice data according to the noise environment presumed based on the noise information. The optimum standard pattern is selected from the plurality of standard patterns that have been set. S in the noise processing unit
With respect to the voice signal with improved / N, its feature vector is calculated in the analysis unit. The collating unit collates the feature vector with the standard pattern selected by the standard pattern selecting unit to recognize the voice.

Therefore, according to the present invention, the S / N of the input voice signal is improved by the pre-processing in the noise processing section, and the non-stationary noise component which is difficult to be superposed on the standard pattern in advance is removed to some extent. It is possible to set. Further, in the collating unit, the standard pattern can be selected according to the noise condition of the input voice signal with improved S / N, so that a high recognition rate can be obtained even in a noisy environment.

According to another structure of the present invention, similarly to the above, the S / N of the signal is improved by first subtracting the estimated noise spectrum from the speech spectrum obtained by frequency analysis of the input speech signal. It On the other hand, in the selection unit, a plurality of types of code vectors and standard patterns created according to a presumed noise environment are prepared in advance.
Here, it is sufficient to store only the code sequence by vector-quantizing the feature vector of the standard pattern in advance, and the storage requires a small memory amount.
The selection unit obtains noise information regarding the residual noise that cannot be completely removed by the noise processing unit, and selects the optimum one from the standard patterns and codebooks stored in advance based on the noise information. On the other hand, in the analysis section, the feature vector is extracted from the speech signal whose S / N has been improved in the noise processing section, and this feature vector is vector quantized in the vector quantization section. The collating unit collates the quantized feature vector with the standard pattern selected by the selecting unit to recognize the voice. Since the distance calculation for matching in the matching unit is limited to the combination of code vectors, the amount of processing can be reduced by using a method such as preparing a calculation table in advance.

Therefore, according to the present invention, not only a high recognition rate can be obtained even in a noisy environment, but also the standard pattern prepared in advance is vector-quantized, so that the storage memory is small, and Also, the distance calculation of the matching unit can be done with a small amount of processing.

[0015]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of a voice recognition apparatus according to the present invention. In the figure, a voice 101 is captured by a voice input unit 102, and an analog voice signal 10
3 is input to the A / D conversion unit 104. The A / D conversion unit 104 converts the audio signal 103 into a digital signal 105 and inputs this to the noise processing unit 106. The noise processing unit 106 spectrally analyzes the digitized input speech 105 and subtracts the estimated noise spectrum from the speech spectrum to remove noise components mixed in the input speech. The speech signal 107 from which the noise has been removed by the noise processing unit 106 is the standard pattern selection unit 1
10 and the analysis unit 108, respectively. The analysis unit 108 calculates, based on the voice signal 107, a voice feature parameter 109 used for recognition determination. The standard pattern selection unit 110 extracts information about the residual noise component that could not be removed by the noise processing unit 106 from the audio signal 107, and optimally fits the noise condition from a plurality of standard patterns stored in advance. A standard pattern 111 is selected. The matching unit 112 performs voice recognition by matching the feature parameter 109 obtained from the input voice with the standard pattern 111 selected by the standard pattern selection unit 110, and outputs a recognition result 113.

Hereinafter, each part will be described in detail.

FIG. 2 shows an example of the configuration of the noise processing unit 106. The noise processing unit 106 removes noise using a method called a spectral subtraction method.
For spectral subtraction, see Boll, “Suppr
ession of AcousticNoise inSpeech Using Spectral Su
btraction ", IEEE Trsns. on Acoustics, Speech, andSi
gnal processing, Vol.ASSP-27, No.2, April 1979, pp.1
Detailed on 13-120. In FIG. 2, 202 is a waveform cutout unit, 204 is a Fourier transform unit, 206 is a silent section detection unit, 208 is a noise spectrum estimation unit, 210 is a subtraction unit, 212 is a Fourier inverse transform unit, and 214 is a waveform synthesis unit.

In the noise processing section of FIG. 2, the input signal 105 digitized by the A / D conversion section 104 is input to the waveform cutting section 202. The waveform cutout unit 202 cuts out a waveform section for analyzing spectrum information from the input signal, and cuts out a section of about several tens ms at regular intervals. The cut section signal waveform 203 is converted into spectrum data by the Fourier transform unit 204.
Here, a fast Fourier transform can be performed by multiplying the cut-out waveform by a window function that has been conventionally used such as a Hamming window, and then embedding zero data before and after it to obtain the data of the factorial factor of 2. High-speed data processing is realized.
The Fourier-transformed spectrum signal 205 is input to the silent section detection unit 206. The silent section detection unit 206 calculates the power of the section spectrum signal, regards the section whose power value is lower than the threshold value for a certain time or longer as a silent section, and outputs the spectrum signal 207 of the section to the noise spectrum estimation unit 208. Several other methods have been proposed as methods for detecting a silent section (speech section), and it is possible to detect a silent section using these methods. The noise spectrum estimation unit 208 estimates the noise spectrum from the inputted spectrum 207 of the non-voice section and stores it. There are several possible methods for estimating the noise spectrum. For example, the estimation is performed by calculating an average spectrum of spectra for several frames. The subtraction unit 210 subtracts the spectrum of the spectrum signal 205 using the noise spectrum 209 estimated by the noise spectrum estimation unit 208. Now, assuming that the spectrum of the input speech is X (f) and the estimated noise spectrum is N (f), the subtraction is expressed by the following equation.

[0020]

[Equation 1]

In Equation 1, the subtraction is performed on the amplitude of the spectrum, but it is also possible to perform the subtraction by using the power spectrum or by adding the phase component. The spectrum 211 from which the noise component has been removed by the subtraction unit 210 is transformed again into the time domain signal 213 by the Fourier inverse transformation unit 212. This signal 213 synthesizes the waveform cut out at the analysis frame period by the waveform synthesizing unit 214 as the voice waveform 107 and outputs it.

FIG. 3 shows another example of the noise processing unit 106, which estimates a noise spectrum using a reference input. In the figure, reference numerals 302 and 308 are waveform cutout sections, 303 and 309 are Fourier transform sections, 306 is a reference input section, 307 is an A / D conversion section, 311 is a subtraction section, 312 is a Fourier inverse transform section, and 313 is waveform synthesis. It is a department.

In the noise processing section of FIG. 3, the input voice 10
As described with reference to FIG. 2, the waveform 5 is converted into the voice spectrum 304 by the waveform cutout unit 302 and the Fourier transform unit 303, and is input to the subtraction unit 311. On the other hand, the noise component 305 is input to the reference input unit 306. At this time, it is desirable that the reference input unit 306 is configured so that the voice is not input as much as possible and only the noise component is selectively input. The noise component input from the reference input unit 306 is digitized by the A / D conversion unit 307 similarly to the input audio signal, and the waveform cutout unit 308 and the Fourier transform unit 30.
After being converted into the noise spectrum signal 310 by 9, the signal is input to the subtraction unit 311. Instead of directly using the input noise component, it is also possible to provide a noise spectrum estimating unit as shown in FIG. 2 and estimate and use the noise spectrum by using the spectrum of several frames, for example. As in the case of FIG. 2, the subtraction unit 311 removes the noise component by subtracting the noise spectrum 310 from the speech spectrum 304,
The voice waveform 107 is obtained via the inverse Fourier transform unit 312 and the waveform synthesis unit 313.

Incidentally, in the case where the Fourier transform is performed on the signals of two channels as shown in FIG. 3, the processing is performed by using the method of simultaneously obtaining the Fourier coefficients of the two signals by utilizing the redundancy of the Fourier transform. The amount can be reduced. Regarding this two-signal simultaneous FFT method, Nakano et al.
It is described in detail in "Analysis of Ship Running Sound by Fast Fourier Transform", Proceedings of Acoustical Society of Japan (Sho 43-11). The outline is given here.

The FFT normally performs a complex operation,
Since the actual signal is a real number, redundancy occurs. Now, assuming that two kinds of input data are Xi, Yi (i = 0 to N-1) and respective Fourier transforms are Ak and Bk (k = 0 to N-1), two-input data Xi, Yi ,

[0026]

[Equation 2]

When the Fourier transform Ck of Zi is obtained, as is clear from the definition of the Fourier transform,

[0028]

[Equation 3]

[0029] On the other hand, if * represents a conjugate complex,

[0030]

[Equation 4]

Because of the relationship

[0032]

[Equation 5]

Therefore,

[0034]

[Equation 6]

[0035]

[Equation 7]

Is obtained. However, Re represents a real part and Im represents an imaginary part. In other words, if Ck is obtained, then A
k and Bk are obtained. The calculation amount required to obtain Ck is Ak,
Alternatively, it is exactly the same as the calculation amount for obtaining Bk, and Ck
Since the calculation amount for obtaining Ak and Bk is negligible as compared with the calculation amount for obtaining Ck, it is possible to simultaneously perform the Fourier transform of two kinds of signals with the calculation amount which is almost the same as the conventional one. ..

FIG. 4 is a diagram for explaining the analysis unit 108 shown in FIG. Although the LPC cepstrum is used as the analysis parameter in this embodiment, other analysis parameters may of course be used.

In FIG. 4, reference numeral 402 denotes a voice section detection unit,
404 is an LPC analysis unit, and 406 is an LPC cepstrum creation unit. The voice section detection unit 402 includes a noise processing unit 10
The speech section is detected from the noise-removed signal 107 output from the signal No. 6, and the signal waveform 403 of the speech section is output. As a method of detecting the voice section, Tawa, Obata, “High-accuracy detection method of voice section in noise”, Proceedings of Acoustical Society of Japan (Sho 62.
Although various methods such as 3) have been proposed, as a simple example, for example, the short-time power of a signal is calculated, and it is determined whether or not the power of a certain threshold value or more continues for a certain time or more. The audio signal 403 is cut out for each fixed section,
It is sent to the LPC analysis unit 404. LPC analysis unit 404
Is an LP using the input section signal 403 for each frame.
The C coefficient 405 is calculated. LPC analysis is a technique that is very commonly used in the field of speech recognition and is described in many documents. For example, Furui, "Digital Audio Processing" Tokai University Press and so on are explained in detail. Here, how to obtain the LPC coefficient will be briefly described. The Pth-order LPC coefficient αj is the interval signal sequence xt (t = 0
~ N) autocorrelation function

[0039]

[Equation 8]

Using

[0041]

[Equation 9]

It can be obtained by solving
Equation 9 can be efficiently solved by Durbin's recursive solution method. LPC cepstrum creation unit 406
Is the LPC coefficient 405 calculated by the LPC analysis unit 404.
Is used to calculate and output the LPC cepstrum 109. LPC cepstrum 109 has LPC coefficient αn (n = 0, n)
Can be obtained from the following recursive expression.

[0043]

[Equation 10]

FIG. 5 shows the standard pattern selection unit 1 shown in FIG.
10 shows a configuration example of 10. In the figure, 502 is a voice section detection unit, 505 is a noise information analysis unit, 506 is a standard pattern storage unit, and 507 is a noise information collation unit. As the voice data used for creating the standard pattern, voice data in which noise is superimposed is used according to the usage environment of the voice recognition device. Alternatively, the created voice data may be noise-removed using the same method as the noise processing unit 106. The noise-removed signal 107 output from the noise processing unit 106 is the voice section detection unit 5
02, the voice section signal 503 and the noise section signal 504
Separated into and. The noise information analysis unit 505 calculates the S / N by means such as taking the ratio of the average power of the voice section signal 503 and the average power of the noise section signal 504. The standard pattern storage unit 506 stores several types of standard patterns created from audio data having different S / N and S / N of the audio data used for the creation. Noise information matching unit 507
Is the S of the noise-processed signal calculated by the noise information analysis unit 505.
Using the value of / N, the standard pattern 111 created using the S / N voice data closest to the S / N of the noise-processed signal from the standard patterns stored in the standard pattern storage unit 506 is output.

Further, the standard pattern selection unit 110 may have the following configuration. The noise information analysis unit 505 of FIG.
Then, the input noise component 504 is analyzed and the analysis parameter is output. The analysis method used here may be the one using the LPC analysis used in FIG. 4, or another analysis method that well expresses the characteristics of the noise component may be used. The standard pattern storage unit 506 stores several types of standard patterns created from voice data on which different types of noises are respectively superimposed, and the characteristic amount of the noise component superimposed on the voice data used for generation. The feature amount of the noise component is analyzed using the same analysis method as that used in the noise information analysis unit 505. The noise information matching unit 507 matches the feature amount of the noise component of the noise-processed signal with the feature amount of the superimposed noise stored in the standard pattern storage unit 506, and based on the matching result, the standard pattern stored in the standard pattern storage unit 506. Among them, the standard pattern 111 created by using the voice on which the noise closest to the noise component of the noise-processed signal is superimposed is selected and output.

Further, the standard pattern selection section 110 can also be used in combination with the above two configurations.
That is, the S / N of the noise-removed signal in the noise information analysis unit 505
Then, the noise information matching unit 507 outputs the optimum standard pattern 111 from the standard pattern storage unit 506 by performing matching using the S / N and the noise feature amount in the noise information matching unit 507.

The collating unit 112 shown in FIG. 1 includes an analysis parameter 109 analyzed by the analysis unit 108 and a standard pattern selection unit 110.
The degree of similarity with the standard pattern 111 is selected by using a method such as DP matching with the standard pattern 111 selected in step 1, and the highest degree of similarity is output as the recognition result 113. D
P matching has become an indispensable technology in the field of speech recognition, and many documents have been published.
Chiba, “Continuous word recognition based on time normalization of speech using dynamic programming”, ASJ Journal 27, 9, pp. 483 to 500 (Showa 46), and Furui, “Digital Speech Processing” mentioned above. Tokai University Press Association is detailed.

As described above, according to the present embodiment, first, the noise level is reduced by the spectral subtraction in the noise processing unit 106, and the unsteady noise such as impact sound which is difficult to superimpose on the standard pattern in advance. Since noise can be removed to some extent, it is not necessary to consider non-stationary noise so much in the subsequent process of the superposed noise method in the standard pattern selection unit 110. In other words, by performing the pre-processing by the spectral subtraction, the nature of the noise and the S
Conditions such as / N are limited. Therefore, the number of standard patterns to be prepared in advance is reduced, the matching accuracy is increased, and the recognition rate in noise is improved.

Next, FIG. 6 shows the configuration of another embodiment of the present invention. In the figure, the same elements as those shown in FIG. 1 are designated by the same reference numerals.

The difference from the configuration of FIG. 1 is that a selection unit 605 is provided instead of the standard pattern selection unit 110, and a vector quantization unit 606 is newly provided. The selection unit 605 analyzes the information of the residual noise components that cannot be completely removed by the noise processing unit 106, and selects the optimum codebook from the plurality of types of codebooks and the plurality of types of standard patterns stored in advance. Select the best standard pattern. The vector quantization unit 606 performs vector quantization on the feature vector of the audio signal analyzed by the analysis unit 108 using the codebook selected by the selection unit 605. The collation unit 112 collates the feature vector of the input speech vector-quantized by the vector quantization unit 606 with the standard pattern selected by the selection unit 605, and outputs the recognition result. Hereinafter, the configuration and operation of the main part of this embodiment will be described in detail.

First, the selection unit 605 will be described with reference to FIG. In FIG. 7, 701 is a section detection unit, 702 is a noise information analysis unit, 703 is a noise information matching unit, 704 is noise information, 705 is a codebook selection unit, 706 is a codebook, 707 is a standard pattern selection unit, and 708 is a standard. It is a pattern. In the selection unit 605, assuming the environment to be used in advance, pre-processing for noise removal is performed on the speech in which the assumed noise is superimposed with an arbitrary S / N using a method similar to that of the noise processing unit 106. Are created as voice data, and codebooks 706-1 to 706-n used for vector quantization are created using these voice data. Vector quantization is a quantization method in which a plurality of parameter sets (vectors) are collectively represented by one code. As shown in FIG. 9, each of the codebooks 706-1 to 706-n has
Various vectors (code vectors) 902 are stored in advance by using a clustering method, and a code (codeword) 901 is associated with each vector. That is, the codebook No. A codebook shows the correspondence between the codeword 901 and the code vector 902 for each 900.

FIG. 10 shows an outline of the operation of the vector quantizer 606. The vector quantizer 606 includes an analyzer 10
The inter-vector distances are respectively calculated with respect to the plurality of code vectors of the selected codebook for the analysis parameter which is the output of one frame of 8, and the code vector with the smallest distance is selected, and the codeword Is output. In the illustrated example, the codeword i is output because the inter-vector distance between the code vector of the codeword i and the code vector of the input analysis parameter is minimum. The vector quantizer 606 repeats this process for each frame.

Note that as a literature on vector quantization, Y. Linde, A. Buzo et al, "An Algorithm for Vecto
r Quantizer Design ”IEEE Trans. on Communications,
vol.COM-28, No.1 Jan.1980, pp.84-95, etc.

Similarly, as shown in FIG. 11, n types (708-1 to 708-n) of standard patterns used for collation are created using n types of voice data. A predetermined number N of N words for each of the n types of voice data
o. The code string is associated with. Each code string is obtained by vector quantization with the corresponding codebook. As described above, by adopting a vector-quantized code string using the codebooks 706-1 to 706-n created by using the same voice data as the standard pattern, the amount of memory required for storing the standard pattern is reduced. can do. Further, by limiting the comparison target standard pattern of the feature vector, the distance calculation in the collation can be efficiently performed. This feature contributes to downsizing of the device and speeding up of processing, and is particularly suitable for in-vehicle and form voice recognition devices.

On the other hand, residual noise information is obtained in advance by using a method similar to that of the voice section detecting unit 701 and the noise information analyzing unit 702 described later. The residual noise information 704-1-7 is shown in FIG.
An example of 04-n is shown. The same figure (a) shows S / as noise information.
N is adopted, and FIG. 7B shows that the power spectrum of the noise component is adopted as the noise information.

As described above, the noise information 704-1 to 704-n, the codebooks 706-1 to 706-n, and the vector quantized standard pattern 708-1 to n are respectively applied to the n kinds of voice data. 708-n is created and stored. Here, the noise information 704-k is for the kth speech data created, similarly, the codebook 706-k is created using the kth speech data, and the standard pattern 708-k is k. It was created using the audio data created in the third step. K at this time
Is called the noise environment number.

The noise removal signal output from the noise processing unit 106 is separated by the voice section detection unit 701 into a voice section signal and a non-voice section (noise section). As a method of detecting the voice section, the same method as the section detection method described in the noise spectrum estimation unit 208 may be used. The noise information analysis unit 702 analyzes and outputs information regarding the residual noise component that cannot be completely removed by the noise processing unit 106 from the output of the voice section detection unit 701. The information to be output is
For example, the S / N ratio of the input voice is obtained using the input voice section signal and the noise section signal, and the S / N ratio is also used. It is also possible to obtain a parameter (power spectrum in the example of FIG. 12B) and use it.

The noise information collating unit 703 collates the residual noise information of the input voice obtained by the noise information analyzing unit 702 with the stored noise information 704-1 to 704-n to make an assumption beforehand. From the n types of noise environments, the one that best matches the current environment is found, and the noise environment number k is output. Based on the information k obtained from the noise matching unit 703, the codebook selecting unit 705 is the most suitable for the current noise environment from the n kinds of codebooks 706-1 to 706-n stored in advance. The codebook 706-k created by using the voice data is selected and output to the vector quantization unit 606. Similarly, the standard pattern selection unit 707 also stores n types of standard patterns 708-stored in advance based on the information k obtained from the noise matching unit 703.
Standard pattern 708-k created by using audio data most suitable for the current noise environment from 1 to 708-n
Is selected and output to the matching unit 112. The selection unit 60
The codebook and the standard pattern in 5 may be selected for each analysis frame, for each unit time, or for each input unit of voice.

Next, the vector quantizer 606 includes a selector 6
Using the codebook 706-k selected in 05, vector quantization is performed on the feature vector of the input speech analyzed by the analysis unit 104, and a quantized code string is output.

The matching unit 112 obtains the degree of similarity between the feature vector of the vector-quantized input voice and the standard pattern selected by the selection unit 605, using the above-described method such as DP matching, with each recognition target, The one with the highest similarity is output as the recognition result. Here, since both the feature vector of the input voice and the feature vector of the standard pattern have already been vector quantized, the distance calculation between the feature vectors is only a combination of code vectors used for vector quantization. Therefore, for each codebook, the distance between code vectors is calculated in advance for each combination, and the calculated distances are stored in a table as shown in FIG. At the time of matching, the distance between the feature vectors can be obtained by referring to the corresponding table, and the amount of matching processing can be greatly reduced. Figure 1
In the example of No. 3, it is immediately known that the distance between codeword 2 and codeword 3 is “3.2”, for example.

If the codebook used for vector quantization of the input speech and the codebook used for vector quantization of the selected standard pattern can be different, a distance table between different codebooks is prepared. You can also Further, instead of preparing a table for each codebook in advance, the table contents may be updated each time the codebook to be used is switched. Further, the distance calculation is performed without preparing a table in advance, the combinations of calculated values and vectors are sequentially stored, and the stored values are used when the distance calculation of the same vector combination is required again. It is also possible to do so.

In this embodiment, the codebook and standard patterns may be selected by using the following configuration.

FIG. 8 is a diagram for explaining a system configuration for selecting a codebook and standard patterns. In FIG. 8, reference numeral 606 is a vector quantization unit, 803 is a quantization evaluation unit, and 804 is a standard pattern selection unit. In the present configuration, the vector quantization unit 606 uses each of the n number of codebooks 802-1 to 802-n prepared in advance to obtain a vector for the feature vector of the input voice obtained by the analysis unit 108. Quantization is performed, and the vector-quantized code string and the quantization error are output. Quantization evaluation unit 80
3 detects the minimum quantization error when using the individual codebooks output from the vector quantization unit 606, and the code vector 802-k used at that time is optimal for the current noise environment. The code string is regarded as a codebook, and the noise environment number k and a vector-quantized feature vector code string are output. In the standard pattern selection section 804, n standard patterns 805-1 to 805 prepared in advance are prepared.
Using the noise environment number k obtained by the quantization evaluation unit 803 from among n, the standard pattern 80 optimal for the current noise environment is obtained.
Select 5-k.

Up to now, it has been explained that the noise environment assumed in advance and the codebook and the standard pattern have a one-to-one correspondence, but which codebook is used for the assumed noise environment, or If the correspondence such as which standard pattern is used is clearly associated, it is not necessary for each to correspond one-to-one. Therefore, the types of codebooks and standard patterns to be stored in advance are assumed. It can be less than the type of noise environment.

For example, in each codebook, if each codebook is created so that the features of the voice components possessed by the code vectors having the same codeword are equal, noise-free voice data can be obtained. It is possible to correspond to each noise environment by preparing one kind of code string data using the standard pattern created by using it and using it in combination with a codebook corresponding to each noise environment. The following method may be used to create such a codebook. When creating a feature vector from voice data, by assigning a number to the analysis frame and referring to that number at the end of clustering, information about from which frame of voice data the feature vector in each cluster is obtained. can get. Therefore, if the feature vector obtained by analyzing the noise-superimposed speech is clustered using the information of this frame to create a codebook, no matter which noise-superimposed speech is used for the codebook, The characteristics of the voice components possessed by the code vectors of the same codeword should be equal. In addition, a codebook is created by using the feature vector obtained from the voice data in which noise is not superimposed,
Next, a similar codebook can be created by superimposing a noise feature vector according to the assumed noise environment on each code vector. As described above, if the common standard pattern is used for all noise environments, the standard pattern selection unit can be omitted in this embodiment.

[0066]

As described above, according to the present invention, the S / N of the input speech is improved by performing the preprocessing by the spectral subtraction on the input speech, and the noise which cannot be removed by the preprocessing is eliminated. On the other hand, the recognition rate of the recognition device can be improved by performing the matching using the standard pattern created from the voice data of the noise-removed signal having a property close to the noise environment.

Further, since the standard pattern prepared in advance is vector-quantized, the storage memory can be reduced, and the distance calculation of the collation unit is limited to the combination of code vectors. It is possible to reduce the processing amount by using a method such as preparing.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a system configuration of an embodiment of the present invention.

FIG. 2 is a block diagram of a configuration example of a noise processing unit in FIG.

FIG. 3 is a block diagram of another configuration example of the noise processing unit in FIG.

FIG. 4 is a block diagram of a configuration example of an analysis unit in FIG.

5 is a block diagram of a configuration example of a standard pattern selection unit in FIG.

FIG. 6 is a block diagram showing a system configuration of a second exemplary embodiment of the present invention.

7 is a block diagram of a configuration example of a selection unit in FIG.

FIG. 8 is a block diagram showing a system configuration example for selecting a codebook and standard patterns.

9 is an explanatory diagram of the contents of the codebook of FIG. 7.

10 is an explanatory diagram of an operation of the vector quantization unit in FIG.

FIG. 11 is an explanatory diagram of the contents of the standard pattern of FIG.

12 is an explanatory diagram of contents of noise information of FIG. 7. FIG.

13 is an explanatory diagram of contents of a table that can be used for collation by the collating unit in FIG.

[Explanation of symbols]

101 ... Input voice, 102 ... Voice input unit, 104 ... A /
D conversion unit, 106 ... Noise processing unit, 108 ... Analysis unit, 11
0 ... Standard pattern selection unit, 112 ... Collation unit, 113 ... Recognition result, 605 ... Selection unit, 606 ... Vector quantization unit.

Claims (21)

[Claims] 1. A voice input section for inputting a voice to be recognized, frequency analysis of an input voice signal obtained by the voice input section, and subtraction of an estimated noise spectrum from the frequency-analyzed voice spectrum. By using a noise processing unit that removes noise, an analysis unit that obtains a feature vector of the input speech based on the speech signal obtained by the noise processing unit, and noise-superimposed speech data that corresponds to a presumed noise environment. A plurality of standard patterns created in advance are stored, and noise information regarding residual noise that cannot be completely removed by noise processing from the voice signal obtained from the noise processing unit is obtained,
A standard pattern selection unit that selects an optimum standard pattern from the plurality of standard patterns based on the noise information, and a standard pattern selected by the standard pattern selection unit and the feature vector are collated and input. A voice recognition device comprising: a collating unit for recognizing voice. 2. The speech recognition apparatus according to claim 1, wherein the subtraction of the spectrum in the noise processing unit is performed using only the amplitude information of the spectrum. 3. The speech recognition apparatus according to claim 1, wherein the subtraction of the spectrum in the noise processing unit is performed using amplitude information and phase information of the spectrum. 4. The speech recognition apparatus according to claim 1, wherein the noise processing section estimates a noise spectrum using an input speech signal obtained from the speech input section. 5. The noise processing section detects a voiceless section from an output signal of the voice input section and estimates a noise spectrum using information of the voiceless section. Speech recognizer. 6. A noise reference input section for referring to noise is provided in addition to the voice input section, and the noise processing section obtains a spectrum obtained from a reference noise signal obtained from the noise reference input section. The speech recognition apparatus according to claim 1, wherein the speech recognition apparatus is used as an estimated noise spectrum. 7. The means for calculating a spectrum in the noise processing section, which performs Fourier transform by regarding one of an input speech signal and a reference noise signal as a complex number signal sequence having a real part and the other as an imaginary part. 7. The speech recognition apparatus according to claim 6, further comprising means for separating the Fourier transform of two signals from the result of the Fourier transform. 8. The standard pattern selection unit has a plurality of standard patterns created from voice data having different S / N, and the S / N from the voice signal obtained by the noise processing unit.
2. The voice recognition device according to claim 1, wherein the standard pattern created from the voice data closest to the S / N value is selected as the standard pattern used in the matching unit. 9. The standard pattern selection unit has a plurality of standard patterns created from voice data on which noises having different properties are superimposed, and a residual noise component is obtained from the voice signal obtained by the noise processing unit. 2. A standard pattern created from voice data obtained by extracting the feature amount of ## EQU1 ## and superimposing noise having a property closest to the property of the residual noise component is selected as a standard pattern used in the matching unit. Voice recognition device. 10. The standard pattern selection unit has a plurality of standard patterns created from voice data in which noises having different properties are superimposed by changing S / N, and the voices obtained by the noise processing unit are included. 2. The S / N and residual noise component feature quantities are calculated from the signal, and the standard pattern created from the voice data closest to these values is selected as the standard pattern used by the matching unit. The voice recognition device described. 11. The standard pattern selecting section uses, as the voice data used for creating the standard pattern, voice data on which noise is superimposed, which is subjected to processing equivalent to that of the noise processing section. The speech recognition device according to Item 8, 9 or 10. 12. A voice input unit for inputting a voice to be recognized, frequency analysis of an input voice signal obtained by the voice input unit, and subtraction of an estimated noise spectrum from the frequency-analyzed voice spectrum. A noise processing unit for removing noise, an analysis unit that obtains a feature vector of the input voice based on a voice signal obtained by the noise processing unit, and a plurality of types of noise generated in advance according to a noise environment assumed. A codebook and a plurality of types of standard patterns as a code sequence obtained by vector quantization using the codebook are stored, and the speech signal obtained from the noise processing section can be completely removed by noise processing. A selection unit that obtains noise information related to residual noise that has not existed and selects the type of the codebook and the type of the standard pattern based on the noise information. , A vector quantizer that obtains a voice feature pattern by performing vector quantization on the feature vector obtained by the analysis unit using the codebook of the selected type; A voice recognition device comprising: a collating unit that collates with a standard pattern of a selected type to recognize the input voice. 13. Instead of using the selecting unit, vector quantization is performed in the vector quantizing unit using each of the codebooks prepared in advance, and the quantization error of the vector quantization is 13. The speech recognition according to claim 12, wherein the smallest one is adopted as a result of vector quantization, and the standard pattern is selected by using information on which codebook is used for the adopted quantization. apparatus. 14. The code book and the standard pattern prepared in advance in plural types are created by using a plurality of audio data having different S / N. Voice recognition device. 15. The code book and the standard pattern prepared in advance in a plurality of types are created using a plurality of audio data on which noises having different properties are superimposed. The voice recognition device described. 16. The code book and the standard pattern prepared in advance in a plurality of types are created by using a plurality of voice data in which noises having different properties are superimposed at various S / Ns. The voice recognition device according to claim 12 or 13. 17. The codebook prepared in advance in a plurality of types is created by superimposing a noise component on a codebook created using speech data on which noise is not superposed. Claim 12 or 1
The voice recognition device according to 3. 18. The voice recognition according to claim 12, wherein, in each of the codebooks prepared in advance, a plurality of code vectors having the same voice feature correspond to the same codeword. apparatus. 19. The method according to claim 12, wherein the code sequences of one standard pattern are made to correspond to different code books, respectively, so as to correspond to various noise environments.
Alternatively, the voice recognition device according to item 13. 20. A table in which the distance between code vectors is calculated in any combination between the codebook used for vector quantization of the input speech and the codebook used for vector quantization of the standard pattern in the matching unit. 13. The voice recognition device according to claim 12, wherein the distance calculation for collating the characteristic pattern of the input voice with the standard pattern is performed by referring to the table. 21. In the distance calculation for matching the voice feature pattern and the standard pattern, the matching unit stores a combination of code vectors calculated once and the distance at that time, and again stores the distance of the same combination. 13. The voice recognition apparatus according to claim 12, wherein the stored value is used when the calculation is performed.