JPH0830294A - Device and method for voice recognition - Google Patents

Abstract

PURPOSE:To provide a voice recognition device which conducts high precision voice recognition of noisy signals by predicting the dynamic featured value of an inputted voice. CONSTITUTION:Only noise is beforehand inputted in order to extract the static featured value of the noise (301). The static featured value of the noise and the static featured value of voice only which is beforehand prepared are added (302) to generate a static featured value code book 304 of a noise mixed voice. Then, differential computations are conducted employing the book 304 and the static featured value of voice only which is beforehand prepared (403) and a conversion table 405 is generated to convert the dynamic featured value of the noise mixed voice into a dynamic featured value of voice only for all featured values of the noise mixed voice. During a voice recognition, an object noise mixed voice is inputted and a voice only static featured value of the input voice is computed (303). On the other hand, the inputted voice is converted into a voice only dynamic featured value by applying the table 405 (404 and 306). Based on the voice only static and dynamic featured values, the recognition processes are conducted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an improvement of a voice recognition device for performing voice recognition by creating a model adapted to noise at any time in an environment in which noise characteristics fluctuate frequently.

[0002]

2. Description of the Related Art A voice mixed with background noise differs from a voice uttered in a noise-free environment in a characteristic parameter used for voice recognition such as a spectrum extracted from the voice. Therefore, in order to maintain a high identification rate when performing speech recognition under noise, it is necessary to perform some noise removal processing or adaptation processing such as model re-creation.

Hidden Markov Model (HMM) of discrete distribution type or Semi-Continuous type which is a standard speech recognition method
In the method of using a codebook such as, for the codebook generated by clustering the features, the model is regenerated with noisy speech and the mapping is used to utter in the noise. Devices that recognize voice are known.

In recent years, a hidden Markov model is separately prepared for each of speech and noise, and MAP (maximum
a posteriori probability) There is an apparatus using a method of recognizing by generating a noisy cepstrum by convolving the cepstrum which is an estimation or speech feature on a linear spectrum.

As for the conventional speech recognition method, for example, B.-H. Juang and LR, Rabiner. “Signal restorat
ion by spectral mapping, ”inProc.IEEE Int.Conf.Aco
ust., Speech, Signal Process., ICASSP87, pp.2368-2371,
Apr.1987, H.Gish, Y.-L.Chow and JRRohlicek, “Prob
abilistic vector mapping of noisy speech parameter
s for HMM word spotting, ”in Proc.IEEE Int.Cnf.Aco
ust., Speech Signal Process., ICASSP90, pp.117-120, Ap
r. 1990, MJ Gales, S. Young, "An Improved approach to
the Hidden Markov Model Decomposition of speech a
nd noise ”, ICASSP92, SFBoll: IEEE Trans.ASSP-27, N
o2 (1978) etc.

[0006]

By the way, as the feature quantity of the voice used for the voice recognition, the input voice is composed of a plurality of frames (continuously repeated sections of a certain time length).
There is a static feature amount obtained by analyzing each of these frames by dividing the static feature amount and a dynamic feature amount obtained by differentiating the static feature amount with respect to time.

However, at the time of speech recognition under noise, the static feature quantity can be estimated by using the above-mentioned method of superimposing the voice feature quantity cepstrum on the linear spectrum. To estimate a dynamic feature such as obtaining a delta cepstrum indicating a dynamic feature by differentiating the cepstrum with time,
An approach different from the above static feature estimation method is required.

Due to such circumstances, the above-described conventional speech recognition method estimates the static feature amount, but does not estimate the dynamic feature amount.

Therefore, even if the static feature amount estimation can deal with the shift of the static feature amount due to the background noise,
It is not possible to deal with the deviation of the dynamic feature amount due to background noise,
It was difficult to further improve the accuracy of voice recognition under noise.

Therefore, an object of the present invention is to provide a voice recognition device capable of performing more accurate voice recognition under noise by estimating a dynamic feature amount of an input voice.

[0011]

According to the present invention, there is provided a device for calculating a static feature amount and a dynamic feature amount from a noise-containing speech to be recognized and passing the result to a speech recognition process. A silent section is input to, and based on the feature quantity of noise included in this silent section and the feature quantity of only voice prepared in advance, the dynamic feature quantity of noise-containing speech is changed to the dynamic feature quantity of only voice. A conversion table creating means for creating a conversion table for conversion, a means for inputting a noise-containing speech to be recognized at the time of speech recognition to obtain a dynamic feature amount thereof, and a dynamic feature amount of the obtained noise-containing speech And a means for estimating the dynamic feature amount of only the voice in the input noise-containing voice by applying the conversion table created in advance, and the estimated dynamic feature amount of the voice only Voices characterized by being passed to recognition processing Identification device is provided.

Further, according to the present invention, there is provided a method for estimating a dynamic feature quantity which is executed by the above speech recognition apparatus.

[0013]

In the present invention, the feature amount of noise included in the silent section is regarded as the estimated feature amount of noise mixed in the input voice, and the feature amount of noise and the feature amount of only voice prepared in advance are considered. Based on the above, a conversion table for converting the dynamic feature amount of the actually input noise-containing voice into the dynamic feature amount of only the voice without noise is created in advance. Then, at the time of voice recognition, the dynamic feature amount is first obtained from the input noise-containing voice, and then the conversion table is applied to this dynamic feature amount to estimate the dynamic feature amount of only the voice. Regarding the static feature amount, the static feature amount of only the voice can be obtained based on the conventional technique. In this way, not only the static feature amount but also the dynamic feature amount, the feature amount of only the voice without the influence of noise is obtained, and as a result, the accuracy of voice recognition is improved.

[0014]

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

First, as a matter related to the present invention, a general method for estimating a static feature amount cepstrum of a voice will be described.

Characteristic obtained by analyzing the noise-containing voice in the process of the voice recognition, using an actual noise-containing voice in which background noise is mixed in the voice to be recognized as an input for the voice recognition. Let the amount of cepstrum be Cs + n.
Here, Cs is the cepstrum of the feature quantity that should be obtained from the same voice that does not contain noise.

At this time, it is considered to maximize the posterior probability density p (Cs | Cs + n), that is, the conditional probability of the event Cs under the condition Cs + n. In the process of this maximization, the cepstrum of the noise component is subtracted from the cepstrum Cs + n of the noisy speech on the linear spectrum, and the difference obtained by this subtraction is converted into the cepstrum again, whereby the cepstrum of the noise-free speech is converted. Cs can be estimated.

However, in the subtraction process, a phenomenon such as a spectral component becoming negative may occur, which is not possible in reality, and thus a large estimation error may occur in the estimation of the noise-free speech cepstrum Cs.

Therefore, by adding target noise to the voice data used for learning for voice recognition, a noise codebook associated with a codebook created with voice containing no noise is created in advance. There is a method of performing speech recognition by mapping (mapping) from this noise codebook to a codebook created with speech that does not contain the noise. Others include noise-free speech cepstrum Cs
There is a method of adding and the cepstrum of the noise component on the linear spectrum and converting again to the cepstrum.

FIG. 1 conceptually shows the relationship between the static feature amount of voice (refers to both voice that does not include noise and voice that includes noise. The same applies hereinafter) and the dynamic feature amount of voice. .

The dynamic feature amount of voice is generally a time-differentiated static feature amount of voice. For voice recognition,
The unit of analysis of the input voice is basically a continuous section having a certain time width called a frame. The input voice is divided into a plurality of frames (frames are shown by the horizontal axis in FIG. 1) in the process of voice recognition, and an analysis process is performed in units of each frame. The static feature amount of the noise-free voice (voice without noise) indicated by 101 is calculated. In the processing process for speech recognition, the dynamic feature amount of voice is obtained by differentiating the static feature amount of voice as described above,
The dynamic feature amount is a value (indicated by reference numeral 102 in FIG. 1) of the inclination of the discrete time series (indicated by the line segment AB in FIG. 1) of the static feature amount of the voice in the past and future frames with respect to the horizontal axis. ) Will be indicated.

When noise is added to the input voice, the influence of noise occurs as indicated by reference numeral 103 in FIG. 1, and the motion of the noise-added voice having a value different from the static feature amount 101 of the noiseless voice is generated. The characteristic amount 104 is obtained. Therefore, the dynamic feature amount of the noise-added voice, which is indicated by the value of the slope of the discrete time series (indicated by the line segment CD in FIG. 1) with respect to the horizontal axis, which is created by the static feature amount of the noise-added voice, is also indicated by reference numeral 105. So it is different from that of noiseless voice 102.

Regarding the static feature amount of speech, there have been attempts to adapt to noise or remove the influence of noise using the above-mentioned various methods. The present invention relates to a method of applying a dynamic feature amount of a voice to an attempt in those static feature amounts of a voice.

FIG. 2 shows the overall construction of an embodiment of a voice recognition device to which the present invention can be applied.

As shown in the figure, this general apparatus has an input terminal 201, an A / D conversion section 202, and a voice section detection section 20.
3, a frame division unit 204, an acoustic analysis unit 205, a voice recognition processing unit 206, an output terminal 207, a feature quantity codebook 208, and a recognition dictionary 209.

Input terminal 2 from a microphone or telephone line
The audio signal (analog signal) input via 01 is
It is quantized by the A / D conversion unit 202. The quantized sample is divided into a voice section and a silent section by the voice section detection unit 203. Further, the quantized sample is divided into a plurality of frames by the frame division unit 204 and output to the acoustic analysis unit 205.

The acoustic analysis unit 205 includes a noiseless static feature quantity codebook (denoted by reference numeral 501 in FIG. 5) and a noiseless dynamic feature quantity codebook (reference numeral 307 in FIG. 3).
, The static feature amount of the voice and the dynamic feature of the voice for each frame output from the frame division unit 204 over both the voice period and the silent period. Extract the amount.

The static feature amount of voice and the dynamic feature amount of voice extracted by the acoustic analysis unit 205 are output to the voice recognition processing unit 206. Both of these feature amounts are subjected to voice recognition processing in the voice recognition processing unit 206 in comparison with the recognition dictionary 209, and the recognition result obtained by this processing is
It will be output from the output terminal 207.

FIG. 3 shows the acoustic analysis unit 2 of this voice recognition device.
The configuration generally adopted as 05 is shown.

As shown in the figure, the general structure of the acoustic analysis unit 205 is a feature quantity extraction section 301, a noise-containing static feature quantity generation section 302, a static feature quantity quantization processing section 303, and a noise-containing static feature quantity. Codebook 304, delta feature quantity extraction unit 30
5, a dynamic feature amount quantization processing unit 306, a noiseless dynamic feature amount codebook 307, and a plurality of delay processing units 308.

In this structure, before extracting the feature quantity of the voice section, first, the frame division unit 2 shown in FIG.
04, the noise section is input through the feature quantity extraction unit 301, and the noise-included static feature quantity generation unit 302 uses the noise-included static feature quantity codebook 30 based on the noise section.
4 is generated in advance and saved.

Next, the speech to be recognized (noise-incorporated speech) is input from the frame division unit 204 in units of one frame, and the feature amount extraction unit 301 performs LPC (LPC: Linear Predictive).
The static feature amount of the recognition target speech is obtained by performing processing such as Coding) analysis. Since this static feature amount is a continuous amount, in order to reduce the amount of calculation in the speech recognition processing unit 206, the static feature amount quantization processing unit 303 performs quantization using the noise-containing static feature amount codebook 304. Converted into static feature amount. The quantized static feature amount is output from the static feature amount quantization processing unit 303 to the voice recognition processing unit 206 illustrated in FIG.

On the other hand, the feature quantity as the continuous quantity is output from the feature quantity extracting section 301 to each delay processing section 308, and subjected to delay processing for several frames in each delay processing section 308 and stored (see FIG. The current frame is stored in the delay processing unit 308 at the middle left end, and the previous frame is stored toward the right delay processing unit 308). The delta feature amount extraction unit 305 calculates the dynamic feature amount of the recognition target voice (noise mixed voice) using the frames output from each delay processing unit 308. This dynamic feature amount is transferred from the delta feature amount extraction unit 305 to the dynamic feature amount quantization processing unit 30.
6 is output. Since this dynamic feature quantity is also a continuous quantity, the dynamic feature quantity quantization processing unit 306 prepares a noiseless dynamic feature quantity codebook 307 (that is,
It is converted into a quantized dynamic feature amount using a dictionary for dynamic feature amounts). This dynamic feature amount is output from the dynamic feature amount quantization processing unit 303 to the speech recognition processing unit 206 in FIG.
In the voice recognition processing unit 206, voice recognition processing is performed together with the static feature amount output from the static feature amount quantization processing unit 303.

FIG. 4 shows the configuration of an embodiment in which the present invention is applied to the acoustic analysis unit 205 of the voice recognition device shown in FIG.

The acoustic analysis unit 205 according to this embodiment is shown in FIG.
In addition to the general configuration shown in FIG. 3, a dynamic feature amount conversion table generation unit 403, a vector conversion unit 404, and a dynamic feature amount conversion table 405 are further provided. By adding these elements, when the dynamic feature amount of the input voice is deviated due to the background noise, the vector conversion unit 404 corrects the shift, thereby dynamically changing the noise-free voice. The feature quantity can be estimated. As a result, it is possible to further improve the accuracy of voice recognition. In FIG. 4, the same components as those shown in FIG. 3 are designated by the same reference numerals.

The vector conversion unit 404 needs the dynamic feature amount conversion table 405 in order to perform vector conversion of the dynamic feature amount of the voice obtained by analyzing the input voice. This dynamic feature amount conversion table 405 is generated by the noise-incorporated static feature amount generation unit 302 subsequent to the generation of the noise-involved static feature amount codebook 304 before the voice input is started. It is generated by the unit 403.

Next, this embodiment will be described in more detail. First, the noise-containing static feature amount generation unit 302 will be described.

FIG. 5 shows the static feature quantity generation unit 302 with noise.
Shows the details of.

As shown in the figure, the noise-containing static feature amount generation unit 302 includes a noiseless static codebook 501 feature amount extraction unit 502, a cosine conversion unit 503, a linear conversion unit 504, an addition unit 505, and a logarithmic conversion unit 506. , Inverse cosine conversion unit 50
7. Feature quantity storage unit 508 and noise spectrum generation unit 50
9 is provided.

First, the noise spectrum generation section 509 analyzes a silent section (a section including only noise) output from the feature amount extraction section 301, such as before or after vocalization of a voice, to obtain a noise power spectrum Φn. It is calculated and used as an estimated value of the noise power spectrum.

Next, only one static feature amount of the noiseless voice stored in the cepstrum state is extracted from the noiseless static feature amount codebook 501 prepared in advance. This static feature is cosine in order to convert from the cepstrum state to the linear power spectrum state.
After the cosine conversion in the conversion unit 503, the linear conversion is performed in the linear conversion unit 504.

Here, the relationship between the cepstrum Cs of noiseless speech and the linear power spectrum Φs of noiseless speech is
It is shown by the following equation (2).

Φs = exp (Cos (Cs)) (1) where Cos () represents cosine transformation.

The linear power spectrum Φs calculated by the above equation (1) is added to the noise power spectrum Φn output from the noise spectrum generation unit 509 in the addition unit 505 to synthesize the spectrum of noise-containing speech. It This spectrum is obtained by the logarithmic conversion unit 50.
In 6 and the inverse cosine transform unit 507, the inverse transform is performed as shown in the following equation (2) and the result is returned to the cepstrum to obtain the estimated cepstrum Cs + n.

[0045]

[Equation 1] The noise-containing cepstrum Cs + n calculated by the equation (2) in the logarithmic transformation unit 506 and the inverse cosine transformation unit 507.
Is registered in the noise-containing static feature quantity codebook 304 in the feature quantity storage unit 508. The above-described processing is repeatedly performed for all the feature amounts in the noise-free static feature amount codebook 501, thereby completing the setting of the noise-containing static feature amount codebook 304.

Subsequently, the dynamic feature quantity conversion table 405 is set as follows.

FIG. 6 shows the details of the dynamic feature amount conversion table generation unit 403.

The dynamic feature quantity conversion table generation unit 403
As shown in the figure, the feature amount extraction units 603 and 604, the conversion vector calculation unit 605, and the conversion vector storage unit 607 are provided.

The generation of the dynamic feature amount conversion table 405 by the dynamic feature amount conversion table generation unit 403 is performed by generating the noise-free static feature amount codebook 501 after the generation of the noise-containing static feature amount codebook 304. This is done by using the noisy static feature codebook 304.

First, the feature quantity (cepstrum of noiseless speech) Cs is extracted from the noiseless static feature quantity codebook 501 prepared in advance by the feature quantity extracting section 603, and the feature quantity corresponding to this feature quantity Cs ( The cepstrum Cs + n of the noise-containing speech is extracted from the noise-containing static feature quantity codebook 304 by the feature quantity extraction unit 604.
At this time, the power spectrum Φs of the feature quantity Cs (power spectrum of noise-free speech) is given by the following equation (3), and the power spectrum Φs + n of the feature quantity Cs + n (power spectrum of noise-containing speech). ) Is given by the following equation (4).

Φs = exp (Cos (Cs)) ……………………………… (3) Φs + n = exp (Cos (Cs + n)) …………………… (4 ) Furthermore, the following equation (5) is derived from the relationship of Φs + n = Φs + Φn.

Exp (Cos (Cs + n)) = exp (Cos (Cs)) + Φn (5) This equation (5) is used for the cepstrum Cs + n of noise-containing speech and the cepstrum Cs of noiseless speech. It can be regarded as the relational expression of. Here, for simplification of notation, x = Cs + n and f (x) = Cs are set in the equation (5), and both sides are partially differentiated by X, and the following equation (6) is derived.

[0053]

[Equation 2] By rearranging this equation (6), the following equation (7) is derived.

[0054]

(Equation 3) However, Cos' (x) is

[Equation 4] Is.

Here, by substituting x = Cs + n and f (x) = Cs and rearranging, the following equation (9) is obtained.

[0056]

(Equation 5) Since cosine transformation is a linear variable, Cos' (x) is constant regardless of x. The same applies to the inverse cosine transform. Therefore, Cos-1 '(Cs) and Cos' (Cs + n) in the equation (9)
Is a constant value (matrix).

Noise-included static feature quantity All feature quantities in the codebook 301 (that is, cepstrum of noise-included speech)
For Cs + n, f ′ (Cs + n) is the conversion vector calculation unit 60.
5 is calculated from the equation (9). And this f '
(Cs + n) is associated with the feature amount Cn + s, and the conversion vector storage unit 606 stores the dynamic feature amount conversion table 405.
Will be stored in.

Next, the vector conversion section 404 will be described.

FIG. 7 shows details of the vector conversion section 404 of FIG.

The vector conversion unit 404 according to this embodiment is
As illustrated, the maximum likelihood static feature amount selection unit 701 and the vector generation unit 702 are provided.

Here, the static feature amount (cepstrum) obtained by analyzing the input voice is Cin. First, in the maximum likelihood static feature amount selection unit 701, the posterior probability density p (Cs + n | Cin) is included in the static feature amount of the quantized speech output from the static feature amount quantization processing unit 303. The quantized feature amount (cepstral) Cs + n that maximizes is selected. The selected feature amount Cs + n is regarded as a static feature amount of noise-containing speech.

The vector generation unit 702 uses the selected feature amount Cs + n and the dynamic feature amount conversion table 405 to generate a vector based on the dynamic feature amount obtained by analyzing the input voice. . The vector generation procedure in the vector generation unit 702 is as follows.

That is, from the definitional expression of f (x), f '(Cs + n) = dC
Since s / dCs + n, and dCs / dt is the dynamic feature quantity stored in the dynamic feature quantity conversion table 405, the above equation (9) is given by the following equations (10) and (11). It is transformed into an expression.

[0064]

(Equation 6) Both ΔCs and ΔCs + n in the equation (11) are dynamic feature quantities. From the equation (11), the dynamic feature amount conversion table 4
It can be seen that 05 is a conversion matrix of the dynamic feature amount (ΔCs, ΔCs + n). Therefore, the dynamic feature amount ΔCs + n extracted from the noisy speech is multiplied by the conversion matrix (conversion matrix in the dynamic feature amount conversion table) described on the right side of Expression (11) to perform vector conversion. , It is possible to estimate the dynamic feature amount ΔCs of only voice (that is, voice without noise).

As described above, according to the embodiment of the present invention, the vector conversion unit 404 selects the maximum likelihood static feature amount Cs + n from the output of the static feature amount quantization processing unit 303. ,
The conversion matrix of the dynamic feature quantity corresponding to the maximum likelihood static feature quantity Cs + n is read from the dynamic feature quantity conversion table 405,
By multiplying this by the dynamic feature amount of the noise mixing unit output from the delta feature amount extracting unit 305, the dynamic feature amount of this noise-containing speech is vector-converted into the dynamic feature amount of only speech that does not contain noise. Therefore, even if the dynamic feature quantity extracted from the delta feature quantity extraction unit 305 is deviated due to the background noise, the deviation does not cause a problem that the accuracy of voice recognition is lowered. Therefore, it is possible to extract the feature amount of the voice with higher accuracy, and it is possible to improve the voice recognition performance.

The above contents are only related to one embodiment of the present invention, and it goes without saying that the present invention is not limited to the above contents.

[0067]

As described above, according to the present invention,
By estimating the dynamic feature amount of the input voice, it is possible to provide a voice recognition device capable of performing more accurate voice recognition in noise.

[Brief description of drawings]

FIG. 1 is an explanatory diagram conceptually showing changes in static voice feature and dynamic voice feature due to background noise.

FIG. 2 is a functional block diagram showing an entire example of a voice recognition device to which the present invention can be applied.

FIG. 3 is a functional block diagram showing a configuration generally adopted in the acoustic analysis unit shown in FIG.

4 is a functional block diagram showing the configuration of an embodiment in which the present invention is applied to the acoustic analysis unit shown in FIG.

FIG. 5 is a functional block diagram showing details of a noise-containing static feature amount generation unit in the embodiment.

FIG. 6 is a functional block diagram showing details of a dynamic feature quantity conversion table generation unit in the embodiment.

FIG. 7 is a functional block diagram showing details of a vector conversion unit of the embodiment.

[Explanation of symbols]

 205 acoustic analysis unit 301 feature amount extraction unit 305 delta feature amount extraction unit 308 delay processing unit 403 dynamic feature amount conversion table generation unit 404 vector conversion unit 405 dynamic feature amount conversion table

Claims (2)

[Claims] 1. An apparatus for obtaining a static feature amount and a dynamic feature amount from a noise-containing speech to be recognized and passing the result to a voice recognition process, in which a silent section is input in advance and included in this silent section. Conversion table creating means for creating a conversion table for converting the dynamic feature quantity of noise-containing speech into the dynamic feature quantity of only voice based on the feature quantity of noise to be generated and the feature quantity of only voice prepared in advance A means for inputting the noise-containing speech to be recognized at the time of voice recognition to obtain a dynamic characteristic amount of the noise-containing speech; Means for estimating a dynamic feature amount of only the voice in the input noise-containing voice by applying the created conversion table, and the estimated dynamic feature amount of the voice only Characterized by passing to recognition processing And a voice recognition device. 2. A method for obtaining a static feature amount and a dynamic feature amount from a noise-mixed speech to be recognized and passing them to a voice recognition process, in which a silent interval is input in advance and included in this silent interval. A process of creating a conversion table for converting a dynamic feature amount of a noise-containing voice into a dynamic feature amount of only a voice based on a feature amount of noise to be generated and a feature amount of a voice prepared in advance; The process of inputting the noise-containing speech to be recognized at the time of recognition and obtaining the dynamic feature amount of the noise-containing voice, and the dynamic feature amount of the obtained noise-containing voice, which has been created in advance A step of estimating a dynamic feature amount of only the voice in the input noise-containing voice by applying a conversion table, and applying the estimated dynamic feature amount of the voice only to the voice recognition process. Speech recognition method characterized by passing Law.