JPH05289695A - Voice recognition system under noise - Google Patents

Abstract

PURPOSE:To efficiently remove a background noise in consideration of fine variation in background noise power and to improve the recognition rate of a voice which is vocalized in noisy environment. CONSTITUTION:The voice recognition system under noise which discriminates the input voice by deforming input voice vectors by using the varying direction of the input voice vectors due to the fine variation in noise power calculated by using spectrum information on the background noise of the input voice and a variation quantity set in advance is provided with a vector movement quantity calculation part 10 which analyzes the noise section preceding the voice section and calculates the variation quantity used for the deformation of the input voice vectors according to the dispersion of the noise power in the noise section, and, the variation quantity is determined according to the analysis of the noise section to perform the voice recognition wherein the time variation quantity of the noise power is accurately reflected.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition system used in noise, and more particularly to a noisy speech recognition system suitable for improving the recognition rate for speech uttered in noise. is there.

[0002]

2. Description of the Related Art A background noise-containing voice, that is, a noisy voice has different characteristic parameters such as a spectrum extracted from a voice uttered in a noise-free environment. Therefore, when recognizing noisy speech,
In order to maintain a high discrimination rate, it is necessary to perform some noise elimination processing or to discriminate in consideration of parameter deformation.

One of the conventional techniques for improving the recognition rate of such noisy speech is, for example, a spectral subtraction method (hereinafter referred to as SS method). This SS
In the method, first, pretreatment such as spectrum analysis described below is performed. In this case, in order to remove the noise component mixed in the voice, by estimating the noise spectrum from the noise-free period preceding the voice period and subtracting it from the spectrum obtained from the noise-containing voice, Get the spectrum. After that, the feature amount of the input voice is obtained, and the identification process is performed with the standard pattern that is set and registered in advance, and the input voice is identified based on whether or not their similarity is within a predetermined threshold value. To do.

Further, as one of the techniques for coping with the change of the characteristic parameter due to the superposition of noise, there is generally a technique of creating a standard pattern by using a voice uttered under background noise in a place where a voice recognition device is used. Are known. For example, in the multi-template method, the signal-to-noise ratio (SN ratio) is
Set several stages, superimpose noise on the voice at the set level, create multiple standard patterns from those noise superimposed signals,
The feature vector extracted from the input voice is the S of the input voice.
The voice under noise is identified by identifying the template with one of the N ratios.

As a conventional technique for improving the recognition rate of noisy speech, for example, S. F. By Boll, "IEEE Trans. ASSP-27, No.
2 (1979) "and" Proceedings of the Acoustical Society of Japan, 1-5-5 (flat 3-10) "by Kasuya et al., The proposer of the present invention.
There is a basis of the present invention described in the above.
In other words, the input speech vector is deformed in consideration of the minute fluctuation of the background noise to perform the classification, thereby reducing the deterioration of the classification rate due to the fluctuation of the noise power. This technology is based on a model in which the spectrum of background noise can be regarded as constant to some extent, and only its power fluctuates slightly, and the effect of noise power fluctuations on the feature vector extracted from speech is taken into consideration. By changing the voice vector, the input is made to follow the power change. The conventional noisy speech recognition technology which is the basis of the present invention will be described with reference to FIG.

FIG. 5 is a block diagram showing the configuration of a conventional noisy speech recognition system. In the figure, 1 is an input device for inputting voice or background noise, 52 is a voice recognition processing device for recognizing input voice, and 3 is
An output device for outputting the recognition result of the voice recognition processing device 52,
Reference numeral 4 is a standard pattern storage device for registering standard patterns used in the recognition processing of the voice recognition processing device 52. The voice recognition processing device 52 responds to the voice input from the input device 1 by
Preprocessing unit 5 for performing preprocessing such as LPC spectrum analysis
And a feature extraction unit 6 for extracting a feature vector from the input speech preprocessed by the preprocessing unit 5 and features of the present speech recognition system under noise, for the feature vector extracted by the feature extraction unit 6, A feature vector transformation unit 7 that performs transformation based on a change in noise power caused by a temporal change in background noise, an input speech vector transformed by the feature vector transformation unit 7, and a standard registered in the standard pattern storage device 4. An identification unit 8 for identifying the input voice based on the distance to the feature vector of the pattern, and a vector change direction calculation unit 9 for calculating the change direction of the vector used in the transformation process of the feature vector transformation unit 7 for the feature vector. There is. The characteristic vector of the standard pattern registered in the standard pattern storage unit 4 is obtained by analyzing noise-free speech in advance. With such a configuration, in the noisy voice recognition system, in the voice recognition processing device 52, the vector change direction calculation unit 9 calculates the moving direction when a minute noise power (background noise power) is added to the input voice. Then, the matching process is performed after the feature vector is provided with redundancy in the moving direction by the variance ratio (σ) indicating the amount of change in the changing direction of the input voice vector that is uniquely determined.

The operation of the voice recognition processing device 52 will be described below. In the speech recognition processing device 52, a linear prediction analysis (LPC: L) is used as a feature amount for speech recognition.
An inner predictive coding) cepstrum or cepstrum is used. In this case, the cepstrum is defined by the following equation (1).

[Equation 1] However, here, s represents a voice signal, and the cepstrum of the expression of Formula 1 is described as Cs in the meaning of the cepstrum of the voice signal s. Further, the cepstrum Cs is a feature amount used as an input speech vector, and in the SS method or the like, it is an estimated value of the cepstrum in which a noise component is subtracted in advance by preprocessing. At this time, the audio signal s is conversely estimated from the cepstrum Cs.

Here, it is assumed that a minute noise Δn is newly added to the voice signal s. This indicates that the noise power added to the voice fluctuates or the power fluctuates due to the noise power estimation error. Due to the fluctuation of the power of noise, s in Expression 1 is replaced with “s + Δn”,
It is expanded as the following formula (2).

[Equation 2] Here, the change in cepstrum due to the small noise Δn is ΔC, and

[Equation 3] Assuming that

[Equation 4] Is obtained, and the ΔC vector calculated from the equation (4) is used as the changing direction of the cepstrum when noise is added, as shown in FIG.

FIG. 6 is an explanatory diagram showing the processing operation relating to the transformation of the input speech vector of the noisy speech recognition system in FIG. This figure shows a modification of the input voice vector by newly using the input voice vector in consideration of the above-described change direction ΔC of the cepstrum. That is, the input voice vector 62 viewed from the reference vector (reference vector) 61 is changed in a direction 64 in which the input voice vector is changed by noise based on a value (variance ratio) σ indicating a predetermined change amount. The resulting vector 63 is regarded as a new input speech vector, shows a large similarity in the direction in which the noise power fluctuates, and is evaluated as a small value in the direction perpendicular to it. The apparent standard pattern is changed as shown by the ellipse in the figure. By using such a distance measure, it is possible to obtain the effect of modifying the input speech vector in consideration of noise power fluctuations.

The technique using the discrimination measure considering such a minute fluctuation of noise power uses the spectrum information of the noise mixed in the input speech in the temporal minute fluctuation of the noise power, and the input speech vector has the noise power of the noise power. The change direction is calculated by a minute change, and the similarity in the change direction is evaluated to be larger than the direction perpendicular to the change direction. Therefore, by changing the input speech vector, the discrimination considering the noise power change is performed. Is a technique for performing. However, this technique uniquely determines the amount of change in the direction of change of the input voice vector. Therefore, it is not possible to accurately reflect the temporal fluctuation amount of noise power.

[0011]

The problem to be solved by the present invention is that the conventional technique uniquely determines the amount of change in the direction of change of the input speech vector, so that the amount of temporal change in noise power is accurately determined. This is a point that cannot be reflected in. The object of the present invention is to solve these problems of the prior art,
It is an object of the present invention to provide a noisy speech recognition system capable of efficiently removing background noise in consideration of minute fluctuations in background noise power and improving the recognition rate for speech uttered under noise.

[0012]

In order to achieve the above object, the noisy speech recognition system of the present invention uses (1) the spectral information of the background noise mixed in the input speech to reduce the noise power due to the background noise. The change direction of the input voice vector due to the change is calculated, and the calculated change direction and the change amount set uniquely are used to deform the input voice vector, and the input voice vector after the change is used, In a noisy speech recognition system that matches the feature vector of the standard pattern and identifies the input speech, the noise section preceding the speech section is analyzed and the input speech vector is transformed based on the variance of the noise power in this noise section. It is characterized in that a vector movement amount calculation unit for calculating a variation amount used for is provided.

[0013]

In the present invention, the vector movement amount calculation unit analyzes the noise section preceding the voice section, obtains the power time series, and calculates the variance value thereof. Then, it is assumed that the fluctuation of the noise power continues even in the voice section, and the calculated variance value (σ) is used as the estimated value to determine the weighting coefficient. In this way, by using the power variance value of the noise section preceding the speech section and setting an appropriate value for redundancy with respect to the input speech vector, it is possible to perform speech recognition in consideration of noise characteristics, It is possible to improve the identification performance of voice mixed with noise as compared with the conventional case.
Further, when the SS method is used, it is possible to further reduce the adverse effect due to excess or deficiency of the noise removal amount.

[0014]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the configuration of the noisy speech recognition system according to the present invention. In the figure, 1 is an input device for inputting voice, 2 is related to the present invention, 3 is a voice recognition processing device for performing a discrimination process of the input voice, 3 is a voice recognition processing device 2.
Output device for outputting the result of voice identification by
Is a standard pattern storage device for storing standard patterns used in the voice recognition processing by the voice recognition processing device 2. still,
The feature vector of the standard pattern registered in the standard pattern storage device 4 is obtained by analyzing noise-free speech in advance. The speech recognition processing device 2 includes a preprocessing unit 5 that performs preprocessing such as LPC spectrum analysis on input speech, a feature extraction unit 6 that extracts a feature vector from the preprocessed input speech, and feature extraction. A feature vector transformation unit 7 that transforms the feature vector extracted by the unit 6 based on a change in noise power caused by a temporal change in background noise, an input speech vector transformed by the feature vector transformation unit 7, and a standard pattern storage device. The vector change direction calculation for calculating the change direction of the vector used in the transformation process for the feature vector transformation unit 7 and the discrimination unit 8 that identifies the input voice based on the distance from the feature vector of the standard pattern registered in FIG. And a vector for calculating the amount of change of the vector used in the transformation process for the feature vector of the feature vector transformation unit 7 according to the present invention. Le movement amount calculating section 1
It is composed of 0 and. Further, the vector movement amount calculation unit 10 includes a noise section detection unit 11 that detects a noise section preceding a speech section, and a power dispersion calculation unit that calculates a noise power dispersion in the noise section detected by the noise section detection unit 11. 12 and a variance ratio determination unit 1 that calculates a variance ratio (σ) indicating the amount of change used to transform the input speech vector based on the variance of the noise power calculated by the power variance calculator 12.
3 and 3.

As shown in FIGS. 2 and 3, which will be described later, there is a correlation between the variance value (σn) of noise power and the spread of the feature quantity. Therefore, in the noisy speech recognition system of this embodiment,
The variance of noise power is used to estimate the spectral variance,
The moving direction and the moving amount when a small noise power (background noise power) is added to the input voice are calculated, and the input voice vector is provided with redundancy for the moving direction and the moving amount, and then the matching process is performed. That is, the preprocessing unit 5
Then, the LPC spectrum analysis is performed on the voice (voice + noise) input from the input device 1, and processing such as the SS method is performed to remove noise to some extent. After that, the feature extraction unit 6 extracts the feature vector. On the other hand, the vector movement amount calculation unit 10 calculates the variance value (σn) of the background noise power, and calculates the input voice vector change amount (variance ratio σ) matching the type of noise. Further, the vector change direction calculation unit 9 obtains the change direction (ΔC) of the cepstrum when a minute background noise is added to the input voice. Then, the feature vector transforming unit 7 transforms the input voice vector based on the change direction (ΔC) calculated by the vector change direction calculating unit 9 and the change amount (variance ratio σ) calculated by the vector moving amount calculating unit 10. Do. Based on the input speech vector obtained as a result, the identification unit 8 performs similarity calculation (matching process) with the feature vector of the standard pattern registered in the standard pattern storage device 4, and outputs the identification result to the output device. Output to 3. In this way, by using the variance of noise power to estimate the spectral variance,
It is possible to set an appropriate value for redundancy with respect to the input voice vector, and it is possible to improve the discrimination performance of the voice in which noise is mixed as compared with the conventional case. Further, when the SS method is used, it is possible to further reduce the adverse effect due to excess or deficiency of the noise removal amount.

FIG. 2 is an explanatory diagram showing an example of temporal power change of background noise input to the speech recognition processing apparatus in FIG. 2A shows an example of changes in the noise power 21 of the station concourse, and FIG. 2B shows an example of changes in the noise power 22 in the telephone box. 2A, 23 is a power dispersion value (σn) calculated from the noise power 21, and FIG. 2B is 24, a power dispersion value (σn) calculated from the noise power 22. σn), which differs depending on the type of noise.

FIG. 3 is an explanatory diagram showing the component distribution of the feature quantity of the speech including each noise in FIG. Figure 3
2A and 2B are respectively shown in FIG.
The noise shown in (a) and (b) is converted into the noise of thousands of samples /
A feature amount is extracted from the voice data superimposed on a /, and the main component analysis is performed and the main component and the second component are displayed. In FIG. 3A, the feature amount spread 31 corresponds to the data indicated by the noise power 21 in FIG. 2A, and in FIG. 3B, the feature amount spread 32 is shown in FIG. Noise power 2 in (b)
It corresponds to the data shown in 2. Thus, as shown in FIGS. 2 and 3, the noise power variance value (σn)
Since there is a correlation between the spread of the feature amount and the spread of the feature amount, it can be said that it is effective to use the variance of the noise power for estimating the spectral variance.

FIG. 4 is a flow chart showing an embodiment of the processing operation according to the present invention of the vector movement amount calculation section in FIG. This embodiment shows a flow of processing for determining an input voice vector change amount from a voice input waveform including noise. First, a noise section is detected from the input speech waveform (step 401), the frame is divided similarly to the speech section, and the noise spectrum and the power are obtained for each frame under the same conditions. From the information, the variance value (σn) of noise power, which is an index value of the degree of noise power variation, is obtained (step 402). Since the transformation of the input speech vector assumes the fluctuation of noise power,
The amount of fluctuation is reflected by analyzing the noise section. In practice, a value proportional to the obtained variance value (σn) is set as the variance ratio (σ) according to the following equation (5) (step 403).

[Equation 5] Here, λ is a constant that is experimentally determined. In this way, by adding the vector movement amount calculation unit that performs this processing to the conventional noise-free speech recognition system shown in FIG. 5, the noise-free speech recognition system shown in FIG. The determined variance ratio (σ) is determined based on the analysis of the noise section preceding the speech section, and the input speech vector change amount can be more accurately determined with respect to the change in the type of input noise. As a result, the identification rate can be improved.

As described above with reference to FIGS. 1 to 4, in the noisy speech recognition system of this embodiment, the variance ratio (σ), which has been uniquely determined in the prior art, is the noise preceding the speech section. The input voice vector change amount can be determined more accurately with respect to the change in the type of the input noise, which is determined based on the analysis of the section. As a result, the identification rate can be improved. The present invention is not limited to the embodiment described with reference to FIGS.

[0020]

According to the present invention, the amount of change in the direction of change of the input speech vector is determined based on the analysis of the noise section preceding the speech section, and the speech recognition accurately reflects the temporal variation of noise power. It is possible to efficiently remove the background noise in consideration of the minute fluctuation of the background noise power, and it is possible to improve the recognition rate for the voice uttered under the noise.

[0021]

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a configuration according to the present invention of a noisy speech recognition system of the present invention.

FIG. 2 is an explanatory diagram showing an example of a temporal power change of background noise input to the voice recognition processing device in FIG.

FIG. 3 is an explanatory diagram showing a component distribution of a feature amount of speech including each noise in FIG.

FIG. 4 is a flowchart showing an embodiment of the processing operation of the vector movement amount calculation unit in FIG. 1 according to the present invention.

FIG. 5 is a block diagram showing a configuration of a conventional noisy speech recognition system.

6 is an explanatory diagram showing a processing operation relating to modification of an input speech vector of the noisy speech recognition system in FIG.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input device 2 Speech recognition processing device 3 Output device 4 Standard pattern storage device 5 Pre-processing unit 6 Feature extraction unit 7 Feature vector transformation unit 8 Identification unit 9 Vector change direction calculation unit 10 Vector movement amount calculation unit 11 Noise interval detection unit 12 Power variance calculation unit 13 Variance ratio determination unit 21, 22 Noise power 23, 24 Power variance value (σn) 31, 32 Feature spread 52 Speech recognition processing device 61 Reference vector (reference vector) 62 Input speech vector 63 Vector 64 Noise The direction in which the input voice vector changes due to

Claims (1)

[Claims] 1. A change direction of an input voice vector when a noise power of the background noise is minutely changed is calculated by using spectrum information of background noise mixed in input voice, and the calculated change direction is uniquely set in advance. The input speech vector is transformed using the set change amount, and the transformed input speech vector is used to perform matching with a feature vector of a standard pattern to identify the input speech. In the above, the noise is characterized by including a vector movement amount calculation means for analyzing a noise section preceding the speech section and calculating a variation used for transforming the input speech vector based on the variance of noise power in the noise section. Lower voice recognition system.