JPH06214596A - Voice recognition device and speaker adaptive method - Google Patents

Abstract

PURPOSE:To obtain a good recognition result by properly compensating for not only individual differences of vocal tract characteristics but also individual differences of vocal cord sound source characteristics and properly adapting uttered voice of a unknown speaker to uttered voice of a standard speaker. CONSTITUTION:Voice recognition process is performed as follows. A frequency characteristic compensating section 10, a frequency axis transformation section 30 and a feature amount extracting section 20 perform process against input voice signals of a unknown speaker with known uttered voice contents for every respective coefficient of plural different frequency characteristic compensating coefficients and plural different frequency axis transformation coefficients, an input voice feature amount is obtained for every coefficient, collating input voice feature amount for every coefficient and standard voice feature amount which has same content of known uttered voice content, select one frequency characteristic compensation coefficient which gives a minimum distance and one frequency axis transformation coefficient among respective coefficients and perform voice recognition processes.

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 a speaker adaptation method used in the field of recognizing voices of unspecified speakers.

[0002]

2. Description of the Related Art Speech has individual differences due to differences in sex, age, physique, voicing method, etc. This individual difference is a major factor that deteriorates the voice recognition performance of an unspecified speaker.
The phoneme-independent individuality includes two variations of the voice spectrum inclination relating to the vocal cord sound source characteristic and expansion / contraction of the voice spectrum relating to the vocal tract characteristic (for example, vocal tract length) in the frequency axis direction. As a method of normalizing such individuality, conventionally, for example, a document “Miwa, Kido:“ Speaker normalization for speech recognition ”(Acoustic Society of Japan, Speech Study Group Material S79-24, July 1979). ) ", Reference" Nakagawa, Kamiya, Sakai: "Recognition of unspecified speaker's word speech based on simultaneous nonlinear expansion / contraction of time axis / frequency axis / intensity axis of speech spectrum" (Journal of the Institute of Electronics and Communication Engineers, Vol. -D, N
o. 2, February, 1981) "and the like, a speaker adaptation method for coping with a pattern variation of a voice signal due to an individual difference has been proposed.

[0003]

However, in the conventional speaker adaptation method as described above, if individual differences are normalized, even some phonological differences are normalized, and phonological properties are lost. There is a problem that there is a large amount of calculation.

In order to solve this problem, Japanese Patent Laid-Open No. 2-25
A speaker adaptation device as disclosed in Japanese Patent No. 9698 has been proposed. In this device, the spectrum of the voice signal of the input speaker is shifted on the frequency axis and converted into the spectrum of the voice signal of the standard speaker. However, speaker adaptation is performed only for shifts on the frequency axis.

By the way, it is desired by those skilled in the art to correct individual differences not only in the voice spectrum relating to the vocal tract characteristics but also in the voice spectrum relating to the vocal cord sound source characteristics without losing the phonological characteristics. In the speaker adaptation device described above, the spectrum relating to the vocal tract characteristics is only normalized on the frequency axis, and the speaker adaptation cannot be performed for the entire speech spectrum with a simple configuration and simple learning. There was a drawback.

According to the present invention, the phonological characteristics are not lost, the device scale is not increased, and the operation is simple and efficient, so that not only individual differences in vocal tract characteristics but also individual differences in vocal cord sound source characteristics can be determined. It is also an object of the present invention to provide a speech recognition apparatus and a speaker adaptation method capable of appropriately correcting the utterance of an unknown speaker to the utterance of a standard speaker, and obtaining a good recognition result. There is.

[0007]

In order to achieve the above object, the present invention has the functions of a speaker adaptation phase and a voice recognition phase, and in the speaker adaptation phase, the known utterance content is unknown. For the input speech signal of the speaker, the frequency characteristic correction means, the frequency axis conversion means, and the feature amount extraction means are processed for each of a plurality of different frequency characteristic correction coefficients and a plurality of different frequency axis conversion coefficients. The input voice feature amount is calculated for each coefficient, and the input voice feature amount for each coefficient is compared with the standard voice feature amount having the same content as the known utterance content, , One frequency characteristic correction coefficient giving the minimum distance and one frequency axis conversion coefficient are selected, and in the speech recognition phase, the unknown speech of the speaker who has input in the speaker adaptation phase is selected. The contents of the input audio signal,
Based on one frequency characteristic correction coefficient and one frequency axis conversion coefficient selected in the speaker adaptation phase, the frequency characteristic correction means, the frequency axis conversion means, and the feature amount extraction means are caused to perform processing to obtain the input voice characteristic amount. Then, the input voice feature amount is collated with the standard voice feature amount held in the standard voice storage means, and the recognition result is output. As a result, it is possible to correct individual differences in vocal cord sound source characteristics and vocal tract characteristics without losing phonological characteristics, and to adapt the utterance of an unknown speaker to the utterance of a standard speaker satisfactorily. It is possible to easily realize speech recognition close to the unspecified speaker speech recognition by using only the standard speech of a person or a small number of standard speakers.

[0008]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an embodiment of a voice recognition device according to the present invention. Referring to FIG. 1, the voice recognition device is
A frequency characteristic correction unit 10 that corrects the frequency characteristic of an input voice signal, a feature amount extraction unit 20 that extracts a cepstrum coefficient of the input voice signal as an input voice feature amount, and a frequency axis conversion is performed on the input voice signal. Frequency axis conversion unit 30
A voice section detection unit 40 that detects a section of the input voice signal, a standard voice storage unit 50 in which the feature amount of the standard voice signal is stored in advance as a standard voice feature amount, and a frequency characteristic with respect to the input voice signal. Correction unit 10, feature amount extraction unit 20,
It has a matching unit 60 for matching (matching) the input voice feature amount obtained by the frequency axis conversion unit 30 with the standard voice feature amount stored in the standard voice storage unit 50.

By the way, in this voice recognition device, the speaker adaptive learning process is performed before the actual voice recognition process is started in order to recognize the voice of an unspecified speaker well. Has become. In order to perform these two types of processing by one device, the device of FIG. 1 further includes a phase selection unit 90 for switching the operation and function of this device to either the speaker adaptation phase or the voice recognition phase. It is provided.

In association with this, the standard voice storage unit 50 stores a standard voice feature amount for speaker adaptation processing and a standard voice feature amount for voice recognition. Further, the frequency characteristic correction unit 10 is prepared in advance with a plurality of different frequency characteristic correction coefficients for speaker adaptive learning, and the frequency axis conversion unit 30 is provided with a plurality of different frequency characteristic correction coefficients for speaker adaptive learning. The frequency axis conversion coefficient of is prepared.

In the speaker adaptation phase, an unknown speaker is made to utter a known utterance content, and the frequency characteristic correction unit 10 and the frequency axis conversion unit 30
The audio signal is processed by sequentially changing the plurality of frequency characteristic correction coefficients and the plurality of frequency axis conversion coefficients, and the matching unit 60 performs the frequency characteristic correction unit 10 and the feature amount extraction in each case. The input voice feature amount obtained by the unit 20 and the frequency axis conversion unit 30 is matched with the standard voice feature amount for speaker adaptation processing stored in the standard voice storage unit 50 to obtain each input voice feature amount and standard voice. The distance from the feature amount is obtained, and the frequency characteristic correction coefficient and the frequency axis conversion coefficient that give the minimum distance are selected and determined.

In the voice recognition phase, the frequency characteristic correction unit 10, for the voice signal of the unknown utterance content of the unknown speaker (actually, the speaker who inputs in the speaker adaptation phase), The frequency axis conversion unit 30 performs processing based on the frequency characteristic correction coefficient and the frequency axis conversion coefficient selected and determined in the speaker adaptation phase, and the matching unit 60 thus performs the frequency characteristic correction unit 1.
0, a standard that gives a minimum distance by matching the input voice feature amount obtained by the feature amount extraction unit 20 and the frequency axis conversion unit 30 with the standard voice feature amount for voice recognition stored in the standard voice storage unit 50. A word corresponding to the voice feature amount is output as a recognition result.

Next, the processing operation of the speech recognition apparatus having such a configuration will be specifically described with reference to the flowcharts of FIGS. 2 and 3 are flowcharts showing processing operations in the speaker adaptation phase and the voice recognition phase, respectively. When the voice recognition device of FIG. 1 is used as a voice recognition device for an unspecified speaker, the user (unknown speaker) adapts his or her voice to the standard voice before performing the actual voice recognition operation. In order to carry out adaptive learning, the phase selector 90 is used to switch the apparatus to the speaker adaptation phase.

In this state, this user (unknown speaker)
Utters five known contents in order, for example, "Hachinohe,""Kesenuma,""Yukuhashi,""Sapporo," and "Kitami," and uses the A / D converter ( Although not shown, can be input via, for example, the sampling frequency 16KH _Z) (step S1). It should be noted that the spectral tilt and frequency shift peculiar to an individual do not depend on the phoneme to the extent that the kind (set) of words for adaptive learning is sufficient to have about 20 syllables. The apparatus of FIG. 1 performs the following processing on these input audio signals.

That is, first, in the voice signal x from the A / D converter, the voice section detector 40 detects the voice section of each word (step S2). Voice section detector 4
At 0, for example, a known two-threshold method using the signal power of each frame and two threshold values is used to detect the section of the input audio signal. It should be noted that the information of this voice section is stored for the processing described later.

The audio signal x from the A / D converter is
The frequency characteristic correction unit 10 inputs the frequency characteristic correction unit 10
Then, the spectral tilt of the input audio signal is corrected by operating a filter H _β (z) represented by the following equation, for example.

[0017]

[Number 1] _{H β (z) = 1-} βz -1 (-1 ≦ β ≦ 1)

The method of correcting the spectral tilt as described above is described in the document "Kano, Sugiyama:" Normalization of the spectral tilt in the LPC spectrum matching scale "(Academic Society of Japan, May 1981, 2-7-15) "
Similar to that shown in. However, in the present embodiment, β is a frequency characteristic correction coefficient, and this is a plurality of values, for example, −0.3, −0.2, −0.1, 0.0, 0.
7 values β _{1 to} β ₇ of 1, 0.2 and 0.3 are taken. As a result, the frequency characteristic correction unit 10 outputs seven types of output signals B _{1 to} B ₇ corrected by the seven types of frequency characteristic correction coefficients β _{1 to} β ₇ , and is added to the feature amount extraction unit 20. The seven types of output signals B _{1 to} B ₇ output from the frequency characteristic correction unit 10 are stored in, for example, a buffer (not shown) for the processing described later.

Each output signal y from the frequency characteristic correction unit 10
_{When 1 to} y ₇ are added, the feature amount extraction unit 20 performs LPC cepstrum analysis on each of the seven types of output signals y _{1 to} y ₇ , and outputs a cepstrum coefficient at constant frame periods. Note that this method is described in "Digital Audio Processing" by Furui (Tokai University Press), September 2, 1985.
5 days ”and the like, and is already known. In particular,
The feature amount extraction unit 20 sets the pre-emphasis: 1-z ^-1 , window period: 16 ms, frame period: 10 ms, LPC analysis order: 14th order, cepstrum analysis order: 14th, LP
C Cepstrum analysis is performed and cepstrum coefficient is output.

Each output signal y from the frequency characteristic correction unit 10
The cepstrum coefficients obtained as a result of the cepstrum analysis for each of _{1 to} y ₇ are sequentially added to the frequency axis conversion unit 30, and in the frequency axis conversion unit 30, these cepstrum coefficients are expressed by, for example, the following equations. The first-order all-pass filter H _α (z) is operated to perform the conversion of the frequency axis, and the conversion result is given to the matching unit 60.

[0021]

H _α (z) = (z ⁻¹ −α) / (1−αz ⁻¹ ) (−1 ≦ α ≦ 1)

The above-mentioned conversion method using the conversion between H _α (z) and z is described in the document “Kobayashi, Matsumoto:“ L ”.
A Study on Frequency Axis Normalization in PC Distance Scale "
(Proceedings of the Acoustical Society of Japan, October 1983, 1-1-
5) ”. However, in the present embodiment, the mel frequency conversion processing in which the mel cepstrum order is, for example, 10th order is also performed at the same time. That is, the value of the frequency axis conversion coefficient α that best approximates the Mel scale is 0.
5 and α is 0.40, 0.4
3, 0.47, 0.50, 0.53, 0.56
Take the seven values α _{1 to} α ₇ of 0.59. These are 0.77 times, 0.84 times, 0.92 times and 1.00 respectively.
Corresponds to frequency axis expansion / contraction of double, 1.09, 1.09, and 1.30.

At this time, in the frequency axis conversion unit 30 and the matching unit 60, first, of the seven types of frequency characteristic correction coefficients β _{1 to} β ₇ , _one frequency characteristic correction coefficient which is most suitable for the input voice of the unknown speaker. Select and determine. Therefore, in the frequency axis conversion unit 30, first, set to ₄ alpha as a reference frequency axis conversion coefficient alpha (= 0.50), by the frequency axis conversion coefficient alpha _4, 7 from the feature extractor 20
The frequency axis conversion of the seed signal is performed, and each conversion result, that is, the mel cepstrum coefficient c _β2 , (in this case,
c ₁₄ , c ₂₄ , ..., C ₇₄ ) are given to the matching unit 60.

In the matching unit 60, the voice section detecting unit 4
The conversion results c ₁₄ , c ₂₄ , ..., C ₇₄ of the input voice signal existing in the voice section detected by 0 from the frequency axis conversion unit 30 are stored in the standard voice storage unit 50 in advance. It is matched (pattern matching) with the standard voice feature amount (specifically, the time series of the mel-cepstral coefficient of the standard voice signal). In addition, this pattern matching is, for example,
Known fixed end point DP (dynamic programming)
Made by law. In that case, the tilt limit of the matching window is
It is "0.5" or more and "2" or less. Also, fixed end point D
In general, the P method has a slight difficulty with respect to a voice section detection error, that is, a difference between the start and end points. However, since the relative comparison between patterns of input voices having the same start and end points is performed here, there is no difference in the start and end points. There is no problem.

In matching by such DP matching, the matching unit 60 selects and saves the coefficient β that minimizes the cumulative distance. Here, the cumulative distance is a total sum of cumulative distances of all five words spoken by the speaker. At that time, it is not necessary to normalize the section length for each word.

In this way, the seven types of characteristic quantities c _{14 to} c ₇₄ of the input audio signal processed by the seven types of frequency characteristic correction coefficients β _{1 to} β ₇ and the reference frequency axis conversion coefficient α ₄ are standardized. The optimum feature amount is selected from the seven types of input voice feature amounts c _{14 to} c ₇₄ by collating with the voice feature amount. When, for example, c ₂₄ is selected as the optimum characteristic amount, the frequency characteristic correction coefficient β ₂ corresponding to this characteristic amount c ₂₄ can be selected and determined as the optimum frequency characteristic correction coefficient (step S3).

Next, when, for example, β ₂ is selected and determined as the optimum frequency characteristic correction coefficient β as described above,
The output signal y ₂ that has already been processed in the frequency characteristic correction unit 10 as described above and stored in the buffer is read from the buffer by this β ₂ and given to the feature amount extraction unit 20. The feature amount extraction unit 20 performs cepstrum analysis on the output signal y ₂ in the same manner as described above, and gives the cepstrum coefficient to the frequency axis conversion unit 30.

The frequency axis conversion unit 30 now selects and determines one frequency axis conversion coefficient most suitable for the input voice of the unknown speaker from the seven kinds of frequency axis conversion coefficients α _{1 to} α _7. Do. For this reason, when the cepstral coefficient for the output signal y ₂ is given from the feature quantity extraction unit 20, the frequency axis conversion unit 30 applies seven kinds of frequency axis conversion coefficients α _{1 to} α ₇ to each of them, thereby making 7 kinds. Frequency axis conversion is performed, and each conversion result, that is, the mel cepstrum coefficient δ _2α (in this case, m ₂₁ , m ₂₂ , ..., m
₂₇ ) is given to the matching unit 60.

In the matching unit 60, the conversion results m ₂₁ , m ₂₂ , from the frequency axis conversion unit 30 are the same as described above.
, M ₂₇ is collated (pattern matching) with the standard speech feature amount of a plurality of utterance contents (specifically, the time series of the cepstral coefficient of the standard speech signal) stored in advance in the standard speech storage unit 50. That is, the matching unit 60 performs DP matching on the input voice feature amount of each set and the corresponding standard voice feature amount (with the same utterance content) for the voice section of each word, and minimizes the cumulative distance. The coefficient α to be selected is selected and saved. Here, the cumulative distance is the sum of the cumulative distances of all five words spoken by the speaker, as described above. At that time, it is not necessary to normalize the section length for each word.

In this way, the seven types of input speech feature quantities m _{21 to} m ₂₇ processed by the one type of frequency characteristic correction coefficient β ₂ and the seven types of frequency axis conversion coefficients α _{1 to} α ₇ are used as standard speech features. 7 types of input speech feature amount m ₂₁
1 to m ₂₇ , one optimal feature amount is selected. When, for example, m ₂₆ is selected as the optimum feature amount, the frequency axis conversion coefficient α ₆ corresponding to this feature amount m ₂₆ can be selected and determined as the optimum frequency axis conversion coefficient, and finally, As the frequency characteristic correction coefficient β and the frequency axis conversion coefficient α that are adapted to the input voice of the unknown speaker who has just uttered,
β ₂ and α ₆ can be determined (step S4).

As described above, in the speaker adaptation phase, the optimum frequency characteristic correction coefficient β and frequency axis conversion coefficient α are selected and determined for the voice of the speaker who wants to recognize his / her own voice. When speaker adaptation is done,
This speaker uses the phase selection unit 90 to switch this device to the voice recognition phase in order to perform the actual voice recognition process.

In the voice recognition phase, this speaker utters unknown contents and causes the device to input the voice (step S11). The input voice of this speaker is the voice section detection unit 4
In addition to 0, the voice section is detected (step S12) and also added to the frequency characteristic correction unit 10. The frequency characteristic correction unit 10 corrects the voice signal of the speaker by the frequency characteristic correction coefficient β, and supplies the output signal to the feature amount extraction unit 20. In the feature quantity extraction unit 20, the frequency characteristic correction unit 10
The cepstrum analysis is performed on the output signal from C to obtain the cepstrum coefficient, and this is given to the frequency axis conversion unit 30. In the frequency axis conversion unit 30, the signal added thereto is frequency axis converted by the frequency axis conversion coefficient α, and the result is matched section 6
Give to 0.

By the way, in the series of processes in the voice recognition phase, the frequency characteristic correction coefficient β and the frequency axis conversion coefficient α which are selected and determined in the speaker adaptation phase are used. That is, in the above example, β ₂ and α ₆ are used as β and α. However, if the speaker adaptation phase has never been performed prior to the voice recognition phase, β and α are set as standard coefficients “0.0”,
“0.5” is used respectively. Therefore, in the voice recognition phase, the input voice signal of the speaker is corrected by the frequency characteristic correction unit 10 by one frequency characteristic correction coefficient β and output as one output signal y.
In the frequency axis conversion unit 30, the frequency axis conversion is performed by one frequency axis conversion coefficient α, and the matching unit 60
, Only one type of input voice feature amount m is sent for this input voice signal.

In the matching unit 60, the voice section detecting unit 4
One kind of conversion result m from the frequency axis conversion unit 30 of the input voice signal existing in the voice section detected by 0 is compared (matched) with the standard voice feature amount stored in the standard voice storage unit 50 in advance. , The standard speech feature amount that minimizes the cumulative distance is selected, and the word corresponding to this is output as the recognition result (step S13).

As described above, in the apparatus of FIG. 1, the learning sample amount for speaker adaptive learning is suppressed as much as possible, and individuals of both vocal cord source characteristics and vocal tract characteristics are maintained without losing phonological characteristics. It is possible to satisfactorily correct the difference and adapt the utterance of an unknown speaker to the utterance of a standard speaker. When applied to the voice recognition of an unspecified speaker, the speech of an unspecified speaker is recognized well. Can be made.

The apparatus shown in FIG. 1 is provided with only one matching section 60, and this matching section 60 is commonly used in the speaker adaptation phase and the voice recognition phase, but as shown in FIG. It is also possible to individually provide dedicated matching units for the speaker adaptation phase and the voice recognition phase.

That is, in the configuration of FIG. 4, the first matching unit 61 that functions only in the speaker adaptation phase,
A second matching unit 62 that functions only in the voice recognition phase is provided, and a first standard voice storage unit 51 is provided for the first matching unit 61 and a second matching unit 62 is provided. The second standard voice storage unit 52
Is provided.

Here, for the first matching unit 61, for example, the same fixed endpoint DP method as described above can be used. In the second matching unit 62, for example, the document “Muroi, Yoneyama:“ Word Speech Recognition Using a Duration Control Type State Transition Model ”(The Institute of Electronics, Information and Communication Engineers Vo1.J
72-D-II, p. 1769, Nov. 1989) ". It should be noted that this speech recognition method has the features that it is more resistant to speech segment detection errors and has higher recognition performance than the fixed endpoint DP method, and that speech recognition can be performed in phoneme units.

As described above, the first matching unit 61 for performing the matching by the matching method suitable for the speaker adaptation processing and the second matching unit 62 for performing the matching by the matching method suitable for the voice recognition processing are separately provided. As a result, the processing in each phase can be performed efficiently and accurately, and a highly accurate recognition result can be obtained.

Further, the apparatus configuration of FIG. 1 can be modified into the configuration shown in FIG. Referring to FIG. 5, unlike the device of FIG. 1, this device is not provided with the phase selection unit 90, and the speaker is not aware of the phase.
The matching unit 60 is configured to perform processing control capable of performing voice input for voice recognition. That is, the matching unit 60 of the apparatus of FIG.
For the voice section, the input voice feature amount and a plurality of standard voice feature amounts are matched (matching), the minimum cumulative distance is obtained from the cumulative distances with each standard voice feature amount, and the minimum cumulative distance is calculated. Only when the distance is less than or equal to a predetermined threshold value, it is assumed that this speaker is uttering a known content (word) for speaker adaptation processing, and the speaker adaptation processing is automatically executed, If it is not less than the predetermined threshold value, it is considered that the speaker uttered in order to recognize the unknown content by voice recognition, and the signal corresponding to the standard voice with the minimum accumulated distance is output as the recognition result, and once. The voice recognition process of is ended.

FIG. 6 is a flow chart for explaining the processing operation of the apparatus of FIG. Referring to FIG. 6, in this device, when an unknown speaker inputs a voice (step S2
1) This voice signal is added to the voice section detection unit 40, and the voice section detection unit 40 detects the voice section from this voice signal (step S22). Further, the input audio signal is subjected to predetermined processing in the frequency characteristic correction unit 10, the feature amount extraction unit 20, and the frequency axis conversion unit 30. That is, in the frequency characteristic correction unit 10, one frequency characteristic correction coefficient currently held therein, for example β ₄ (= 0.
0) corrects the frequency characteristic of the input audio signal, and the frequency axis conversion unit 30 uses the one frequency axis conversion coefficient currently held therein, for example, α ₄ (= 0.5), to extract the feature quantity extraction unit. The frequency axis of the feature amount output from 20 is converted, and the feature amount as the conversion result is given to the matching unit 60. In the matching unit 60, the feature amounts from the frequency axis conversion unit 30 of the input voice signal existing in the voice segment detected by the voice segment detection unit 40 are stored in the standard voice storage unit 50 in various standard voice feature amounts. (Matching), and the minimum cumulative distance is extracted from the cumulative distances with each standard voice feature amount (step S23).

Next, the matching unit 60 determines whether the minimum cumulative distance thus extracted is smaller than a predetermined threshold value. As a result, when the minimum cumulative distance is not smaller than the predetermined threshold value, it is judged that this speaker has uttered an unknown content for voice recognition, and the word corresponding to the standard voice feature amount giving the minimum cumulative distance is determined. Is output as a recognition result, and one speech recognition process is ended (step S
25). On the other hand, when the minimum cumulative distance is smaller than the predetermined threshold value, it is determined that this speaker has uttered a known content (word) for speaker adaptation processing, and the matching unit 60
Causes the frequency characteristic correction unit 10, the feature amount extraction unit 20, and the frequency axis conversion unit 30 to perform the speaker adaptation processing similar to the processing of FIG. 2 (steps S26 and S27).

As described above, the apparatus of FIG. 5 executes the speaker adaptation processing depending on whether or not the minimum cumulative distance obtained as a result of performing the speech recognition processing once is smaller than a predetermined threshold value. Since it is determined whether or not the phase is to be determined, the operator does not have to select the phase and the speaker adaptation process can be automatically executed when it is necessary.

In the apparatus of FIG. 5, the criterion of whether the minimum cumulative distance is smaller than a predetermined threshold is used as the criterion for determining whether or not the speaker adaptation process is to be executed. For example, it is also possible to use the ratio of the distance between each of the first and second candidates for recognition. Also, the coefficient α,
Since β can be a continuous quantity, the coefficients α and β can be smoothed as shown in the following equation when the coefficients α and β are updated in the speaker adaptation phase.

[0045]

Equation 3] _{α t = aα + (1-} a) α t-1, β t = aβ + (1-a) β t-1 (0 ≦ a ≦ 1)

Here, α _t and β _t are frequency axis conversion coefficients and frequency characteristic correction coefficients which are newly set after execution of the present speaker adaptation phase, and α _t-1 and β _t-1 are present speaker adaptation. The frequency axis conversion coefficient that was set before the phase was executed,
Frequency characteristic correction coefficients, α and β are frequency axis conversion coefficients selected in this speaker adaptation phase, frequency characteristic correction coefficients, and a is a constant for smoothing, or for smoothing according to the recognition distance. Is a variable of. Here, when a is used as a variable, as the reliability of the recognition result is higher, the coefficients α and β for speaker adaptation are changed more quickly. Try to get a close value. For example, the cumulative distance may be set to D (> 0), and a may be expressed by the following equation.

[0047]

## EQU00004 ## a = exp (-D)

In each of the above embodiments, the frequency characteristic correction unit 1
The processing is performed in the order of 0, the feature quantity extraction unit 20, and the frequency axis conversion unit 30, but this processing order is not essential, and is described above in the frequency characteristic correction processing and the frequency axis conversion processing. It can vary when other methods different from the method are used. Further, it is also possible to use correction equations and conversion equations used in the frequency characteristic correction unit 10 and the frequency axis conversion unit 30 other than those of the equations 1 and 2, and the number and values of the coefficients are not limited to this example. ,
Any desired number and value can be used depending on the case.

As for the order of coefficient selection in the speaker adaptation phase, the coefficient α may be selected first and then the coefficient β may be selected according to the signal storage capacity or the like, or the processing may be performed. It may be executed simultaneously for all combinations of the coefficients α and β depending on the capability. Also, as for the type (set) of words of utterance content for speaker application, another type (set) of words different from the above example may be used as long as there is no extreme phoneme bias. Further, as for the voice section detection method and the voice recognition method in the voice section detection unit 40 and the matching unit 60, methods other than those described above can be used.

[0050]

As described above, according to the inventions of claims 1 and 2, the functions of the speaker adaptation phase and the voice recognition phase are provided, and in the speaker adaptation phase, known utterances are produced. Frequency characteristic correction means, frequency axis conversion means, feature amount extraction means for each of a plurality of different frequency characteristic correction coefficients and a plurality of different frequency axis conversion coefficients for an input voice signal of a speaker whose content is unknown. The input speech feature amount is calculated for each coefficient, and the input speech feature amount for each coefficient is compared with the standard speech feature amount having the same known utterance content and each coefficient is calculated. From among the above, one frequency characteristic correction coefficient that gives the minimum distance and one frequency axis conversion coefficient are selected, and in the speech recognition phase, the unknown utterance of the speaker who has input in the speaker adaptation phase is selected. The input voice signal of 1) is processed by the frequency characteristic correction means, the frequency axis conversion means, and the feature amount extraction means based on one frequency characteristic correction coefficient and one frequency axis conversion coefficient selected in the speaker adaptation phase. Then, the input voice feature amount is calculated, the input voice feature amount is collated with the standard voice feature amount held in the standard voice storage means, and the recognition result is output.
It is possible to correct individual differences in vocal cord sound source characteristics and vocal tract characteristics without losing phonological characteristics and to adapt the utterance of an unknown speaker to the utterance of a standard speaker satisfactorily. It is possible to easily realize speech recognition close to the unspecified speaker speech recognition by using only the standard speeches of a large number of standard speakers.

According to the third aspect of the present invention, the matching means is suitable for the speaker adaptation processing by using the standard speech feature quantity stored in the first standard speech storage means in the speaker adaptation phase. A first matching means for performing matching by a matching method, and a matching method suitable for speech recognition processing using the standard speech feature amount stored in the second standard speech storage means in the speech recognition phase. Since it is divided into two collating means,
The speaker adaptation process and the voice recognition process can be efficiently and accurately performed.

Further, according to the invention described in claim 4, processing selecting means for switching and selecting between the speaker adaptation phase and the voice recognition phase is further provided, and the collation means is
Since the speaker adaptation phase and the voice recognition phase are switched according to an instruction from the process selection means, the speaker can be started at any time when the speaker starts voice recognition input or when the recognition rate is low. Can easily perform speaker adaptation processing, so-called supervised learning.

According to the fifth aspect of the present invention, the collating means uses the predetermined one frequency characteristic correction coefficient and the predetermined one frequency axis correction coefficient to correct the frequency characteristic, the frequency axis converting means, and the characteristic. The input voice feature amount obtained by processing by the amount extraction means is collated with the standard voice feature amount held in the standard voice storage means, and the input voice signal is a voice signal having a known utterance content based on the comparison result. The speaker adaptation phase is selected when it is judged whether or not there is an input voice signal of known utterance content, so that the speaker input from the voice input whose content is unknown for voice recognition. The adaptation process can be automatically started, the speaker can save the labor of selecting the speaker adaptation phase in advance, and so-called unsupervised learning can be performed.

According to the sixth aspect of the present invention, the matching means obtains the distance between the input voice feature quantity and the standard voice feature quantity as a matching result, and only when the distance satisfies a predetermined condition. Since the speaker adaptation phase is selected, unsupervised speaker adaptation processing can be performed only when the recognition result is highly reliable, and the speaker adaptation processing is inappropriately executed. Can be prevented.

Further, according to the invention of claim 7, there is provided a coefficient smoothing means for smoothing the frequency characteristic correction coefficient and / or the frequency characteristic correction coefficient and / or the frequency axis conversion coefficient in the frequency axis converting means. The coefficient smoothing means is further provided, and when the speaker adaptation phase is selected and executed, the value of the frequency characteristic correction coefficient and / or the frequency axis conversion coefficient immediately before execution of the speaker adaptation phase and the speaker Since the frequency characteristic correction coefficient and / or the frequency axis conversion coefficient are smoothed by using the value after the execution of the adaptation phase, the speaker adaptation processing can be stably executed.

According to the eighth aspect of the present invention, the coefficient smoothing means is a smoothing coefficient determined according to the distance between the standard speech feature quantity and the input speech feature quantity corresponding to the collation result of the speech recognition processing. Since the frequency characteristic correction coefficient and / or the frequency axis conversion coefficient is smoothed by using, when the reliability of the recognition result is low, the smoothing coefficient hardly changes and whether the speaker adaptation process is performed or not is performed. Does not require judgment processing.

According to the invention of claim 9, correction of individual differences in vocal cord sound source characteristics and correction of individual differences in vocal tract characteristics are performed serially, so that speaker adaptation processing can be efficiently performed. it can.

According to the tenth aspect of the invention, the learning sample amount for speaker adaptive learning is a predetermined number (for example, about 20 syllables), and is about 20 syllables sufficient for adaptive learning. In order to obtain utterances, one can easily select a simple and easily vocabulary word set.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an embodiment of a voice recognition device according to the present invention.

FIG. 2 is a flowchart for explaining a processing operation in a speaker adaptation phase of the voice recognition device in FIG.

FIG. 3 is a flowchart for explaining a processing operation in a voice recognition phase of the voice recognition device in FIG.

FIG. 4 is a diagram showing a modification of the voice recognition device shown in FIG.

5 is a diagram showing a modification of the voice recognition device shown in FIG.

FIG. 6 is a flowchart for explaining the processing operation of the voice recognition device in FIG.

[Description of Reference Signs] 10 frequency characteristic correction unit 20 feature amount extraction unit 30 frequency axis conversion unit 40 voice section detection unit 50 standard voice storage unit 51 first standard voice storage unit 52 second standard voice storage unit 60 matching unit 61 1st matching part 62 2nd matching part 90 Phase selection part

Claims (10)

[Claims] 1. A frequency characteristic correction means for correcting the frequency characteristic of an input audio signal based on a plurality of predetermined different frequency characteristic correction coefficients, and a plurality of different predetermined frequency axis conversion coefficients. A frequency axis conversion means for converting the frequency of the input voice signal, a feature amount extraction means for extracting the feature amount of the input voice signal as an input voice feature amount, and a standard holding a standard voice feature amount. Collation means for collating the voice storage means, the frequency characteristic correction means, the frequency axis conversion means, and the input voice feature quantity obtained by processing by the feature quantity extraction means with the standard voice feature quantity held in the standard voice storage means. And a voice recognition device having functions of a speaker adaptation phase and a voice recognition phase, wherein the collating means is configured to detect a known utterance content in the speaker adaptation phase. For an input voice signal of a known speaker, a frequency characteristic correction means, a frequency axis conversion means, and a feature amount extraction means are provided for each of the plurality of different frequency characteristic correction coefficients and the plurality of different frequency axis conversion coefficients. To process the input voice feature quantity for each coefficient,
The input voice characteristic amount for each coefficient is collated with the standard voice characteristic amount having the same content as the known utterance content, and one frequency characteristic correction coefficient and one frequency axis that give the minimum distance from the respective coefficients. Selecting a conversion coefficient, and the collating means, in the voice recognition phase,
Based on one frequency characteristic correction coefficient and one frequency axis conversion coefficient selected in the speaker adaptation phase, with respect to the input voice signal of unknown utterance content of the speaker input in the speaker adaptation phase The frequency characteristic correcting means, the frequency axis converting means, and the feature quantity extracting means are caused to perform an input voice feature quantity, and the input voice feature quantity is collated with the standard voice feature quantity held in the standard voice storing means. hand,
A voice recognition device characterized by outputting a recognition result. 2. The voice recognition device according to claim 1, wherein
The standard voice storage means holds a standard voice feature amount for the speaker adaptation phase and a standard voice feature amount for the voice recognition phase. A speech recognition apparatus characterized in that a standard speech feature amount having the same content as a known content uttered by a speaker in a person adaptation phase is set. 3. The voice recognition device according to claim 1, wherein
The standard voice storage means includes a first standard voice storage means that holds a standard voice feature quantity for a speaker adaptation phase and a second standard voice storage means that holds a standard voice feature quantity for a voice recognition phase. Further, the matching means is configured to perform matching by a matching method suitable for the speaker adaptation process using the standard speech feature amount held in the first standard speech storage means in the speaker adaptation phase. First to do
And the second matching means for performing matching by a matching method suitable for the speech recognition process using the standard speech feature amount held in the second standard speech storage means in the speech recognition phase. A voice recognition device characterized by being configured. 4. The voice recognition device according to claim 1,
Process selection means for switching and selecting between the speaker adaptation phase and the voice recognition phase is further provided,
The collation unit receives an instruction from the process selection unit,
A voice recognition device characterized in that it switches between a speaker adaptation phase and a voice recognition phase. 5. The voice recognition device according to claim 1,
The matching means uses the predetermined one frequency characteristic correction coefficient and the predetermined one frequency axis correction coefficient to process the frequency characteristic correction means, the frequency axis conversion means, and the feature quantity extraction means. Is compared with the standard voice feature amount held in the standard voice storage means, it is determined whether or not the input voice signal is a voice signal of a known utterance content based on the comparison result, A voice recognition device, characterized in that a speaker adaptation phase is selected when it is judged that it is an input voice signal. 6. The voice recognition device according to claim 5,
The matching means obtains the distance between the input voice feature quantity and the standard voice feature quantity as a match result, and selects the speaker adaptation phase only when the distance satisfies a predetermined condition. A voice recognition device. 7. The voice recognition device according to claim 5, wherein
Coefficient smoothing means for smoothing the frequency characteristic correction coefficient and / or the frequency characteristic correction coefficient and / or the frequency axis conversion coefficient in the frequency axis conversion means is further provided, and the coefficient smoothing means When the speaker adaptation phase is selected and executed, the frequency characteristic correction coefficient is used by using the value of the frequency characteristic correction coefficient and / or the frequency axis conversion coefficient immediately before the speaker adaptation phase is executed and the value after the speaker adaptation phase is executed. And / or a frequency axis conversion coefficient is smoothed. 8. The voice recognition device according to claim 7,
The coefficient smoothing means uses the frequency characteristic correction coefficient and / or the frequency axis by using a smoothing coefficient determined according to a distance between a standard speech feature amount corresponding to a matching result of a speech recognition process and an input speech feature amount. A speech recognition device characterized by smoothing conversion coefficients. 9. When selecting one frequency characteristic correction coefficient and one frequency axis conversion coefficient suitable for the voice of an unknown speaker from a plurality of different frequency characteristic correction coefficients and a plurality of different frequency axis conversion coefficients. , With respect to an input voice signal of an unknown speaker of known content, for each of a plurality of coefficients of either one of a frequency characteristic correction coefficient and a frequency axis conversion coefficient, the input voice feature amount for each coefficient and the known Obtain the distance between the utterance content and the standard voice feature amount of the same content, select the coefficient that gives the minimum distance, and then select the input voice feature amount for each coefficient and the known utterance for each of the other types of coefficients. A speaker adaptation method characterized in that a distance between a content and a standard speech feature having the same content is obtained, and a coefficient giving a minimum distance is selected. 10. The speaker adaptation method according to claim 9, wherein the known utterance content uttered by an unknown speaker is:
A speaker adaptation method characterized in that a predetermined number of plural words are used.