JPH07239696A - Voice recognition device - Google Patents

Abstract

PURPOSE:To correct the change in inclination of a spectrum in a section of a voiced sound with a relatively larger sound power and easy to occur voice deformation by performing inversion filtering only in a sound section decided as the voiced sound by an adaptive first inversion filter part. CONSTITUTION:The sound data divided at every fixed interval are inputted to a self correlation function calculation part 201. The calculation part 201 calculates the self correlation function of an input sound. When the input sound is a voiced sound, a large peak should exist in a part equivalent to a repeat period in the autocorrelation function. Then, the maximum value among the peaks of the autocorrelation function excepting a 0-order peak is detected by a peak detection part 202. A decision part 203 compares the maximum value of the peaks obtained by the peak detection part 202 with a threshold value awaited beforehand, and decides that the inputted sound is the voiced sound when the maximum value exceeds the threshold value, and the change in the inclination of the spectrum due to the speaking deformation is corrected by the succeeding inversion filtering in the section of the voiced sound.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus, and more particularly to a noise resistant speech recognition apparatus which operates stably even in the presence of 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 capable of correctly recognizing even a voice uttered under noise.

As a factor of deterioration of recognition performance in noisy speech recognition, there is an influence on noise superimposed on speech. In the field of speech recognition, the spectral subtraction method is known as a very effective method for reducing the influence on the superimposed noise. This method removes noise components by subtracting the estimated noise spectrum from the short-time spectrum of input speech.

However, in addition to the effect of superposed noise, it is known that the stress caused by speaking in a noisy environment causes variations in the speaking style (speaking deformation), which adversely affects recognition. As a measure against voicing deformation, a method of performing speaker adaptation by regarding the voicing deformation voice as the voice of a different speaker,
It has been proposed that the method for normalizing the feature amount of the vocalized modified voice is used.

One of the phenomena of voicing deformation is that the high frequency component of the voice is emphasized and the slope of the spectrum changes. As a method of normalizing the slope of this spectrum, application of adaptive first-order inverse filtering can be considered. In this method, the frequency characteristic is flattened by adaptively performing inverse filtering using a linear prediction coefficient corresponding to the slope of the spectrum.

[0006]

The adaptive first-order inverse filter is effective for improving recognition performance by correcting a change in the slope of the spectrum for voicing modified speech in which noise is not mixed. However, since noise is superimposed on the voice uttered in an actual environment, an inverse filter is applied to the noise component in a voiceless section such as a pause or a section where the voice level is low such as a consonant. It has a negative effect on cognition. For example, many noises such as automobile running noises are concentrated in the low frequency band. When an inverse filter for flattening the frequency characteristic is applied to such noise, the high frequency noise component is emphasized. Therefore, the adaptive first-order inverse filtering applied to speech uttered under noise is
It is necessary to do so without affecting the noise component.

[0007]

In order to achieve the above object, the voice recognition device of the present invention is obtained from a voice input unit for inputting a voice to be recognized and a voice signal input to the voice input unit. An adaptive first-order inverse filter unit that performs inverse filtering on a voice signal using a first-order Percoll coefficient, an analysis unit that calculates a feature vector of the inverse-filtered voice signal, a standard pattern registered in advance, and the analysis. When a matching unit that recognizes the input voice by determining the similarity to the feature vector obtained by the unit and a determination unit that determines the characteristics of the input signal are provided, and the input signal matches the conditions of the determination unit Only has a means for performing inverse filtering in the adaptive first-order inverse filter section.

[0008]

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

In a speech recognition apparatus having an adaptive first-order inverse filter section for performing inverse filtering for flattening frequency characteristics, a voiced / unvoiced decision section for determining whether the input voice is voiced or unvoiced is provided. Inverse filtering is performed only for the voice section determined to be. By this process, the power of the voice is relatively large,
In a voiced sound section that is likely to be affected by voicing deformation, a change in spectral slope due to voicing deformation can be corrected by inverse filtering. On the other hand, by omitting the inverse filtering in the unvoiced section, which has a relatively small voice power and is easily affected by noise, it is possible to avoid adverse effects such as emphasizing a noise component by filtering.

Therefore, according to the present invention, it is possible to improve the recognition performance of the voice accompanied by the deformation of the voice uttered in a noisy environment.

[0011]

EXAMPLES Examples of the present invention will be shown below.

FIG. 1 is a block diagram of a voice recognition system for explaining an embodiment of the present invention. In FIG.
101 is a voice input unit, 102 is a voiced / unvoiced determination unit, 10
3 is an adaptive first-order inverse filter unit, 104 is an analysis unit, 105 is a matching unit, 106 is a standard pattern storage unit, and 107 is a switch unit. In FIG. 1, only the outline of the present embodiment will be described, and a detailed description of each part will be given from FIG. 2 onward. Voice input unit 101
The voice input to is converted into a digital signal by A / D conversion and then divided at regular intervals (usually several tens of ms).
(Analysis frame). The voice data divided for each analysis frame is input to the voiced / unvoiced determination unit 102, and it is determined whether or not it is a voiced sound. Here, when it is determined that the input voice frame data is voiced sound, the adaptive first-order inverse filter unit 103 performs adaptive first-order inverse filtering to flatten the frequency characteristic of the voice data. The adaptive first-order inverse filter will be described in detail later. Also,
When it is determined that the input voice frame data is not voiced sound, the process of the adaptive first-order inverse filter unit 103 is omitted.
Next, the analysis unit 104 calculates a characteristic parameter from the input voice divided for each frame. Standard pattern storage 10
6 stores a standard pattern (feature vector sequence) of the recognition target vocabulary calculated in advance. Of course, the standard pattern stored here calculates the feature vector by the same analysis system as this system. Collation unit 1
Reference numeral 05 performs distance calculation between the standard pattern stored in the standard pattern storage unit 106 and the feature vector of the input voice analyzed by the voice analysis unit 104. At this time, it is determined that the word having the smallest distance among the standard patterns matched by the matching unit 105 is the recognition word of the input voice, and is output as the recognition result.

Next, each processing unit briefly described in FIG. 1 will be described in detail.

First, the voiced / unvoiced determination unit 102 will be described. FIG. 2 is a diagram illustrating an example of the voiced / unvoiced determination unit 102. In FIG. 2, 201 is an autocorrelation function calculation unit, 202 is a peak detection unit, and 203 is a determination unit. The voiced sound is determined by the periodicity of the input voice data. The voice data divided at regular intervals is input to the autocorrelation function calculation unit 201. The autocorrelation function calculation unit 201 calculates the autocorrelation function of the input speech x (i) by the processing shown in Expression 1.

[0015]

[Equation 1]

FIG. 3 shows an example of the autocorrelation function calculated from actual voice data. If the input voice is a voiced sound, there should be a large peak in the portion corresponding to the repetition period (pitch) in the autocorrelation function. Therefore, the peak detection unit 202 detects the maximum value among the peaks of the autocorrelation function excluding the 0th-order peak (see FIG. 3).
In the example, the value of the 68th sample). The determination unit 203 compares the maximum value of the peak obtained by the peak detection unit 202 with a threshold value prepared in advance, and when the maximum value of the peak exceeds the threshold value, the input voice is voiced. Judge that there is.

FIG. 4 is a diagram for explaining a second embodiment of the voiced / unvoiced decision unit 102. In the embodiment described with reference to FIG. 2, the peak value of the autocorrelation function of the input speech is used for the voiced / unvoiced determination, whereas in the embodiment of FIG. 4, the modified correlation function of the input speech (prediction of the linear prediction analysis is used. The peak value of the residual autocorrelation coefficient) is used. In FIG. 4, 401 is a linear prediction analysis unit, 402 is a linear prediction inverse filter unit, 403 is an autocorrelation function calculation unit, 404 is a peak detection unit, and 405 is a determination unit. The voice data divided at regular intervals is input to the linear prediction analysis unit 401. The linear prediction analysis unit 401 performs a linear prediction analysis on the input voice data and outputs a linear prediction coefficient. Linear predictive analysis is a very common analysis method in the field of speech signal processing, and there are many good books that explain Furui's "Digital Speech Processing" in detail. Several algorithms have been proposed for calculating the linear prediction coefficient, but as an example, Levins
The processing flow of the on-Durbin algorithm is shown in FIG. In the linear prediction inverse filter unit 402, the linear prediction coefficient a (i) obtained by the linear prediction analysis unit 401 is applied to the input speech.
Inverse filter using. The inverse filter is performed by the process shown in Formula 2.

[0018]

[Equation 2]

The output ε (n) of the filtering corresponds to the prediction error of the linear prediction analysis and is called the residual. The residual calculated by the linear prediction inverse filter unit 402 is the autocorrelation function calculation unit 40.
Enter in 3. The autocorrelation function calculator 403 calculates the autocorrelation function of the prediction residual. The correlation function calculated here is called the modified correlation function of the input speech. When the modified correlation function is calculated, the peak detection unit 404 detects the maximum value from the peaks of the modified correlation function excluding the 0th-order peak. The determination unit 405 compares the maximum value of the peak obtained by the peak detection unit 404 with a threshold value prepared in advance, and when the maximum value of the peak exceeds the threshold value, the input voice is voiced. Judge that there is.

FIG. 6 is a diagram for explaining a third embodiment of the voiced / unvoiced determination unit 102. Like the first and second embodiments, the present embodiment is also a method for determining voiced / unvoiced based on the periodicity of data. Here, the peak value of the high-keflency component of the cepstrum is used to determine the periodicity. In FIG. 6, 601 is a cepstrum calculation unit, 602 is a peak detection unit, and 603 is a determination unit. The voice data divided at regular intervals is input to the cepstrum calculation unit 601. The cepstrum calculator 601 uses the input voice data
After performing FFT and transforming into the frequency domain and taking the logarithm, I
The cepstrum is calculated by transforming into the time domain again by FFT. FIG. 7 shows an example of the cepstrum calculated from actual voice data. The abscissa of the cepstrum is called keflenency, and the components of the spectrum envelope are concentrated in the low kefflenency portion, and the fundamental frequency is obtained from the peak of the high kefflenency portion. The peak detection unit 602 detects the peak value of this high-keflency portion from the obtained cepstrum. The determination unit 603 compares the peak value obtained by the peak detection unit 604 with a threshold value prepared in advance, and if the peak value exceeds the threshold value, the input voice is a voiced sound. judge.

In addition to the above, there are many methods for judging voiced or unvoiced, and it goes without saying that those methods can also be applied to this embodiment.

Next, the adaptive first-order inverse filter unit 103 will be described. FIG. 8 is a diagram for explaining an example of the adaptive first-order inverse filter unit 103. In FIG. 8, 80
Reference numeral 1 is a first-order Percoll coefficient calculation unit, and 802 is an inverse filter unit. The primary PARCOR coefficient calculating unit 801 calculates the primary PARCOR coefficient k1 of the input voice data from the equation (3).

[0023]

[Equation 3]

Here, r0 and r1 are the 0th order term and the 1st order term of the autocorrelation function, respectively. The inverse filter unit 802 filters the input voice data by using the first-order Percoll coefficient calculated by the first-order Percoll coefficient calculation unit 801. Inverse filtering using the first-order Percoll coefficient has the function of flattening the spectrum, and in the field of speech recognition, it has been reported in the past that it is effective in compensating for spectrum tilt due to high-frequency loss in telephone lines and individual differences. There is. The adaptive first-order inverse filter is formulated by the following equation, where x (n) is the input signal and k1 is the first-order Percoll coefficient.

[0025]

[Equation 4]

By the way, Equation 5 is obtained from Equation 4 and Equation 1.

[0027]

[Equation 5]

By using the equation 5, the autocorrelation function can be directly filtered, and the required processing amount can be reduced as compared with the case where the waveform signal is directly filtered.

The effect of the adaptive first-order inverse filter will be described with reference to FIG. In FIG. 9, 901 is a spectrum of a normally uttered voice, and 902 is a spectrum of a voicing-deformed voice. Reference numeral 903 denotes a spectrum obtained by subjecting a normally uttered voice to an adaptive first-order inverse filtering process, and 904 denotes a spectrum obtained after subjecting a uttered voice to an adaptive first-order inverse filtering process. 901 and 9
Comparing with 02, the power of the high frequency component of the spectrum of 902 is higher than that of 901, and a large difference is seen between the two spectra. On the other hand, in the spectra 903 and 904 of the voice subjected to the adaptive first-order inverse filter processing, it can be seen that the difference between the two is small. That is, it is possible to correct the influence of vocal deformation by using the adaptive first-order inverse filter.

Next, the analysis unit 104 will be described. The analysis unit 104 calculates the characteristic parameter of the voice used when the matching unit 105 calculates the distance from the input voice. There are many characteristic parameters used in speech recognition, such as LPC cepstrum, mel cepstrum, bandpass filter output, and FFT spectrum. In this embodiment, the case of using the most commonly used LPC cepstrum will be described. FIG. 10 is a block diagram for explaining an example of the analysis unit 104. In FIG. 10, 10
Reference numeral 01 is a linear prediction analysis unit, and 1002 is a cepstrum calculation unit. For the speech data input to the linear prediction analysis unit 1001, a linear prediction coefficient (LPC coefficient) is obtained according to the analysis processing flow shown in FIG. Cepstrum calculator 10
02 calculates the LPC cepstrum (c1, ..., cn) from the LPC coefficients (a1, ..., an) by the recursive formula shown in Formula 6.

[0031]

[Equation 6]

Finally, the collating unit 105 will be described.
FIG. 11 is a diagram for explaining the matching unit 105. Figure 1
In FIG. 1, 1101 is a DP matching unit and 1102 is a minimum distance determining unit. The DP matching unit 1101 stores the characteristic parameter (nth-order LPC cepstrum in this embodiment) sequence for each input voice frame obtained by the analysis unit 104 and the standard pattern (feature vector sequence of registered voice) stored in the standard pattern storage unit 106. ) And the distance calculation. Of course, in the creation of the standard pattern, the analysis parameters were obtained after the first-order adaptive inverse filter processing was applied only to the voiced section as in the input speech. DP matching is DTW (Dynamic Ti
Also known as me Warping), normalization for variations in the utterance duration of speech patterns is performed by dynamic programming.
g), which has long been used to recognize isolated words. For details of DP matching, see Furui; "Digital Audio Processing" (Tokai University Press). When the distance calculation with all the standard patterns is completed in the DP matching unit 1101, the minimum distance determination unit 11
02 finds the standard pattern with the smallest distance calculation value. Minimum distance determination unit 11 in the voice recognition system
The registered word of the standard pattern with the minimum distance obtained in 02 is set as the recognition result.

As described above, according to this embodiment,
In a voiced sound section, which has a relatively large voice power and is easily affected by voicing deformation, a change in the slope of the spectrum due to voicing deformation can be corrected by inverse filtering. on the other hand,
By omitting the inverse filtering in the unvoiced section having a relatively small voice power and being easily influenced by noise, it is possible to avoid adverse effects such as emphasizing a noise component by filtering.

Therefore, according to the present invention, it is possible to improve the recognition performance of the voice accompanied by the deformation of the voice uttered in the noisy environment.

As a second embodiment of the present invention, instead of judging voiced / unvoiced and switching the presence / absence of inverse filter processing, it is judged whether the input voice frame is a vowel and the filtering processing is performed. Consider switching between presence and absence. FIG. 12 is a system block diagram for explaining the second embodiment of the present invention. In FIG. 12, 1201 is a voice input unit, 1202 is a vowel determination unit, 1203 is an adaptive first-order inverse filter unit, 1204 is an analysis unit, 1205 is a matching unit, 1206 is a standard pattern storage unit, and 1207 is a switch unit. As in the first embodiment, the voice input unit 1201
The voice input to is converted into a digital signal by A / D conversion and then divided at regular intervals. The voice data divided for each analysis frame is input to the vowel determination unit 1202, and it is determined whether or not it is a vowel. When it is determined that the input voice frame data is a vowel, the adaptive first-order inverse filter unit 1203 performs adaptive first-order inverse filtering to flatten the frequency characteristic of the voice data. When it is determined that the input voice frame data is not a vowel, the processing of the adaptive first-order inverse filter unit 1203 is omitted. Next, the analysis unit 1204 calculates characteristic parameters from the input voice. The standard pattern storage unit 1206 stores standard patterns of the vocabulary to be recognized, which have been calculated in advance. The collating unit 1205 compares the standard pattern stored in the standard pattern storage unit 1206 with the voice analysis unit 1.
Distance calculation is performed with the feature vector of the input voice analyzed in 204. At this time, it is determined that the word having the smallest distance among the standard patterns matched by the matching unit 1205 is the recognition word of the input voice, and the recognition result is output.

Next, each processing unit will be described in detail. By the way, the voice input unit 1201 and the adaptive first-order inverse filter unit 1
The description of 203, the analysis unit 1204, the collation unit 1205, and the standard pattern storage unit 1206 overlaps with the description in the first embodiment. Therefore, the description thereof is omitted, and the vowel determination unit 1
Only 202 will be described.

The vowel section has a feature that it has a pitch and a frame having a relatively large power continues for a fixed time (about 60 ms). In the present embodiment, a vowel determination method based on the magnitude of power will be described as an example. FIG. 13 is a diagram showing an example of the vowel determination unit 1202. In FIG. 13, 1301 is a power calculation unit and 1302 is a determination unit.
The power calculator 1301 calculates the short-time power of the analysis frame of the input voice. In this embodiment, the zero-order term of the autocorrelation function is used as the short-time power. Here, if the autocorrelation function is calculated, it is not necessary for the adaptive first-order inverse filter unit 1203 and the analysis unit 1204 to calculate the autocorrelation function again. The determination unit 1302 prepares a threshold value for the voice power in advance, and when the input voice power exceeds the threshold value for a certain number of consecutive frames, the section is determined to be a vowel section.

Of course, other than this, there are many methods for determining the vowel segment, and it goes without saying that these methods can also be applied to this embodiment.

As described above, according to the second embodiment, in the vowel section in which the voice power is relatively large and the influence of the voicing deformation is likely to occur, the inclination change of the spectrum due to the voicing deformation is corrected by inverse filtering. be able to.
On the other hand, it is possible to avoid adverse effects such as emphasizing the noise component by filtering by omitting the inverse filter processing for the sections (consonant sections, silent sections) other than vowels, which have relatively low voice power and are easily affected by noise. .

Therefore, according to the present invention, it is possible to improve the recognition performance of the voice accompanied by the deformation of the voice uttered in the noisy environment.

The description so far has been made on the assumption that a voice-transformed voice is input as the input voice. Vocal deformation is a problem that occurs only when vocalizing in a high noise environment, and does not occur in a quiet environment. Therefore, as a third embodiment, a method of switching the presence / absence of adaptive inverse filter processing according to the measured ambient noise level will be described. FIG. 14 is a system block diagram for explaining the third embodiment of the present invention. In FIG. 14, 1401 is a voice input unit, 1402 is a noise level measuring unit, 1403 is a noise judging unit, 1404 is a switching unit, 1405 is an adaptive first-order inverse filter unit, 1406 is an analyzing unit, 1407 is a standard pattern storing unit, and 1408 is The standard pattern selection unit 1409 is a collation unit. Voice input unit 140
The voice signal input from 1 is converted into a digital signal by A / D conversion, and then divided at regular intervals (usually several tens of ms) (analysis frame). The noise level measuring unit 1402 measures the noise level of the input data divided for each analysis frame. The noise level measuring unit 1402 will be described later. The noise determining unit 1403 determines the magnitude of the noise of the input signal from the noise level obtained by the noise level measuring unit 1402. That is, the noise level measuring unit 1
When the noise level obtained in 402 is larger than the threshold value, it is determined that the noise is large. Switch part 1404
Switches processing depending on the noise level. If the noise is large, the process is moved to the adaptive first-order inverse filter unit 1405. On the contrary, if the noise is low, the processing is moved to the analysis unit 1406. The adaptive first-order inverse filter unit 1405 performs adaptive first-order inverse filtering to flatten the frequency characteristic of audio data. Also, the analysis unit 1406 calculates a characteristic parameter from the input voice divided for each frame. The details of the adaptive first-order inverse filter unit and the analysis unit have already been described. The standard pattern storage unit 1407 stores standard patterns of recognition target words. In the case of the present embodiment, two types are stored: a standard pattern analyzed through the adaptive first-order inverse filter process and a standard pattern analyzed without performing the adaptive first-order inverse filter process. The standard pattern selection unit 1408 selects a standard pattern according to the magnitude of the noise determined by the noise determination unit 1403. That is, when the noise is large, the standard pattern analyzed through the adaptive first-order inverse filter processing is used, and when the noise is small, the standard pattern analyzed without performing the adaptive first-order inverse filter processing is used. Collator 14
09 performs distance calculation between the standard pattern selected by the standard pattern selection unit 1408 and the feature vector of the input voice analyzed by the analysis unit 1406. At this time, the collating unit 1
It is determined that the word having the smallest distance among the standard patterns checked in 409 is the recognition word of the input voice.

Here, the noise level measuring section 1402 will be described in detail. FIG. 15 shows the noise level measuring unit 140
It is a figure for demonstrating No. 2. In FIG. 15, 150
Reference numeral 1 is a voice section detection unit, and 1502 is a noise power calculation unit. The voice section detection unit 1501 detects a voice section from the input signal. The voice section detection is explained in detail, such as Furui's "Digital Speech Processing". As a general example, a voice section is determined on the basis of a section in which short-time power of a certain threshold value or more continues for a certain time or more. The noise power calculation unit 1502 takes an average of short-time power calculated for each frame. This averaging process continues until the voice section detection unit 1501 detects a voice section. By this processing, the noise level measurement is completed when the voice section is detected.

As described above, according to the third embodiment, it is possible to correct the inclination change of the spectrum due to the utterance deformation by inverse filtering only when the utterance is made in a high noise environment where the utterance deformation is likely to occur. On the other hand, the adverse effect of the inverse filtering process can be avoided by omitting the inverse filtering process in a quiet environment in which no vocal transformation occurs.

Therefore, according to the present invention, it is possible to improve the performance of recognizing a voice accompanied by a vocalization deformation in a noisy environment without causing performance degradation when used in a quiet environment.

Further, the third embodiment and the first embodiment,
A combination with the second embodiment is also possible. For example, in FIG.
In the fourth embodiment shown by, the voiced / unvoiced determination unit 1610 is added to the third embodiment. According to this embodiment, in a voiced sound section in which voice power is relatively large and the influence of vocalization deformation is likely to occur, it is possible to correct the inclination change of the spectrum due to vocalization deformation by inverse filtering. On the other hand, by omitting the inverse filtering in the unvoiced section, which has a relatively small voice power and is easily affected by noise, it is possible to avoid adverse effects such as emphasizing a noise component by filtering. Further, in a quiet environment where voicing transformation does not occur, the adverse effect of the inverse filtering process can be avoided by omitting the inverse filtering process in all sections.

Therefore, according to the present invention, it is possible to improve the recognition performance of a voice accompanied by a voicing deformation in a noisy environment without causing performance degradation when used in a quiet environment.

Of course, it goes without saying that the combined use of the third embodiment and the vowel judging section is also effective.

[0048]

As described above, according to the present invention, in a voiced sound section in which voice power is relatively large and the influence of voicing deformation is likely to occur, it is possible to inversely filter a change in spectrum slope due to voicing deformation. Can be corrected. On the other hand, by omitting the inverse filtering in the unvoiced section, which has a relatively small voice power and is easily affected by noise, it is possible to avoid adverse effects such as emphasizing a noise component by filtering.

Therefore, according to the present invention, it is possible to improve the performance of recognizing a voice produced in a noisy environment and accompanied by a vocal deformation.

[Brief description of drawings]

FIG. 1 is a block diagram for explaining a first embodiment of the present invention.

FIG. 2 is a block diagram for explaining an example of a voiced / unvoiced determination unit.

FIG. 3 is a diagram showing an example of an autocorrelation function of voice data.

FIG. 4 is a block diagram for explaining a second embodiment of a voiced / unvoiced determination unit.

FIG. 5 is a processing flow for explaining an example of a linear prediction analysis unit.

FIG. 6 is a block diagram illustrating a third embodiment of a voiced / unvoiced determination unit.

FIG. 7 is a diagram showing an example of a cepstrum calculated from audio data.

FIG. 8 is a diagram for explaining an example of an adaptive first-order inverse filter unit.

FIG. 9 is a diagram for explaining the effect of an adaptive first-order inverse filter.

FIG. 10 is a diagram for explaining an example of an analysis unit.

FIG. 11 is a diagram illustrating an example of a matching unit.

FIG. 12 is a diagram for explaining the second embodiment of the present invention.

FIG. 13 is a diagram illustrating an example of a vowel determination unit.

FIG. 14 is a diagram for explaining the third embodiment of the present invention.

FIG. 15 is a diagram for explaining a noise level measuring unit.

FIG. 16 is a diagram for explaining the fourth embodiment of the present invention.

[Explanation of symbols]

101 ... Voice input unit, 102 ... Voiced / unvoiced determination unit, 10
3 ... Adaptive first-order inverse filter unit, 104 ... Analysis unit, 105 ...
Collating unit 106 ... Standard pattern storage unit 107 ... Switch unit.

Claims (9)

[Claims] 1. A voice input unit for inputting a voice to be recognized, and an adaptive first-order inverse filter for performing inverse filtering on the voice signal using a first-order Percoll coefficient obtained from the voice signal input to the voice input unit. Section, an analysis section for calculating a feature vector of the inversely filtered voice signal, and recognition of the input voice by obtaining the similarity between the standard pattern registered in advance and the feature vector obtained by the analysis section. In a voice recognition device having a collating unit for performing,
A voice recognition device characterized by comprising a judging section for judging characteristics of an input signal, wherein the adaptive primary inverse filter section carries out inverse filtering only when the input signal matches the condition of the judging section. 2. The adaptive primary inverse function is provided only when a noise level measuring unit for measuring the magnitude of noise mixed in an input voice signal is provided, and the determining unit determines that the noise level exceeds a threshold value. The voice recognition device according to claim 1, wherein inverse filtering is performed by the filter unit. 3. A voiced / unvoiced determination unit for determining whether the input voice is voiced or unvoiced as the determination unit, and the adaptive first-order inverse filter unit performs inverse filtering only on a voice section determined to be voiced sound. The voice recognition device according to claim 1. 4. The voice recognition apparatus according to claim 3, wherein the voiced / unvoiced determination unit determines voiced or unvoiced voice using a peak of an autocorrelation function calculated from the input signal. 5. The voiced / unvoiced determination unit determines voiced or unvoiced voice using a peak of a modified correlation function (automatic correlation coefficient of prediction residual of linear prediction analysis) calculated from the input signal. The voice recognition device according to claim 3, characterized in that 6. The voiced / unvoiced determination unit is characterized by determining voiced / unvoiced voice using a high-keflency portion of a cepstrum calculated from the input signal.
The voice recognition device described. 7. A vowel determination unit for determining the input voice as a vowel section and a section other than the vowel section is provided as the determination section, and the adaptive first-order inverse filter section performs inverse filtering only on a voice section determined to be a vowel. The voice recognition device according to claim 1, characterized in that. 8. The voice recognition apparatus according to claim 7, wherein the vowel discrimination section uses the short-time power value calculated from the input signal to determine the vowel section. 9. An autocorrelation calculation unit for calculating an autocorrelation function of an input speech signal is provided, and the adaptive first-order inverse filter unit performs inverse filtering on the autocorrelation function calculated by the autocorrelation calculation unit. Claim 1 characterized by
8. The voice recognition device according to 8 above.