JPH04281497A - Voice recognition method under noisy environment - Google Patents

Abstract

PURPOSE:To provide a voice recognition method under a noisy environment by preventing the reduction of recognition percentage caused by noise power estimation errors due to time varying noise and providing input vector variation considering a small variation of noise power in a voice recognition system having a multitemplate method and an SS method as the main pattern matching scheme. CONSTITUTION:By adding the variation assuming a slight variation of noise power per 104 against the voice input vectors obtained under the influence of a noisy background, a recognition is performed considering noise power variations. For example, in case of the multitemplate method, variational direction of the input vectors due to the variation of noise is obtained employing a background noise spectrum information and a matching process is performed using the recognition scale which is set to have a redundancy in the direction of the variation.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] The present invention relates to a speech recognition method for a speech recognition device that is used under noise and has a recognition system based on pattern matching, and in particular, the present invention relates to a speech recognition method that is used under noise and has a recognition system based on pattern matching. Concerning speech recognition methods.

0002

2. Description of the Related Art Speech mixed with noise differs from speech produced in a noise-free environment in characteristic parameters such as spectra extracted therefrom. Therefore, when recognizing speech in noise, it is necessary to perform some kind of noise removal processing.
Alternatively, it is necessary to perform identification in consideration of parameter deformation. For example, the spectral subtraction method (hereinafter referred to as
When performing speech recognition using the SS method (abbreviated as SS method), as shown in Figure 2, preprocessing such as spectrum analysis is first performed (201
). In this case, in order to remove the noise component mixed into the speech, the noise spectrum is estimated from the section of only noise without speech, and subtracted from the spectrum obtained from the speech mixed with noise. After this, the features of the input voice are calculated (2
02), performs matching processing with standard patterns that have been set and registered in advance (203), and identifies input audio based on whether or not the distance values are within a predetermined threshold (2).
04). Additionally, as a method for the recognition system to respond to changes in feature parameters due to superimposition of noise, it is generally known that a standard pattern is created using speech uttered under background noise at the location where the speech recognition device is used. It is being For example, in the multi-template method, the signal-to-noise ratio (SN ratio)
Set several levels and superimpose noise on the audio at the set level,
Create multiple standard patterns from those noisy signals,
The feature vector extracted from the input voice is set to the SN of the input voice.
Speech under noise is identified by matching a template that is close to the ratio. For conventional speech recognition methods, see, for example, "Handbook of Electronics, Information and Communication Engineers, edited by the Institute of Electronics, Information and Communication Engineers (1989), pp. 119.
1 to 1207".

0003

[Problem to be solved by the invention] In the above conventional technology, S
When using the S method, in the process of calculating the difference between the (speech + noise) spectrum and the noise spectrum, it may not be possible to completely remove the noise component due to noise estimation errors, or on the contrary, excessive subtraction may be performed. There is a problem that this may have a negative effect on speech recognition. Further, in the multi-template method, problems arise as the number of templates increases. In other words, during recognition, a process is required to compare the feature vector obtained from the input speech with a standard pattern, but by setting multiple S/N ratios, there are many standard patterns for each identification category, which reduces processing time. It takes. Furthermore, if the input voice does not have a template with the same SN ratio, it will be identified using a template that is close to it, so there is a risk that the voice will be erroneously recognized as being in a different identification category. Furthermore, resetting an appropriate template significantly increases processing time. These problems are caused by noise power estimation errors caused by temporal changes in background noise. The purpose of the present invention is to improve such problems and to reduce the recognition rate caused by fluctuations in noise power by performing classification that takes into account minute fluctuations in background noise when performing speech recognition based on pattern matching. An object of the present invention is to provide a method for recognizing speech in noise that can reduce degradation.

0004

[Means for Solving the Problems] In order to achieve the above object, the speech recognition method under noise of the present invention is applicable to a speech recognition method for a system having a recognition system based on pattern matching, such as the multi-template method or the SS method. , using spectral information of the noise mixed in the input audio, calculate the direction in which the input vector changes due to minute fluctuations in noise power, and weight it so that the distance in the direction of change is smaller than the direction perpendicular to the direction of change. A feature of this method is that by setting a distance measure, the input vector is transformed in the direction of change in noise power, and identification is performed using the transformed input vector in consideration of noise power fluctuations.

[0005]

[Operation] In the present invention, when using the multi-template method, the discriminator applies transformations to the input vector of speech input under background noise, assuming minute fluctuations in the background noise. That is, the direction in which the reference vector of the input vector changes is determined using the spectrum information of the background noise, and the distance measure is weighted so that it takes a small value in the direction of change and a large value in the direction perpendicular to it. By using this weighted distance measure, the input vector has redundancy with respect to the direction of change, thereby improving the identification performance when identifying noise-containing speech compared to the conventional method. Note that if the recognition rate is set to be the same as before, the number of templates can be reduced. Furthermore, when using the SS method,
It is possible to reduce the negative effects caused by excess or deficiency in the amount of noise removed.

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. First, the principle of the speech recognition method in noise in this embodiment will be described. FIG. 3 is an explanatory diagram showing the weighting of input vectors in the speech recognition method under noise according to an embodiment of the present invention, and FIG. 4 is an explanatory diagram showing the change direction of input vectors in the speech recognition method under noise according to an embodiment of the present invention. It is a diagram. In this example, linear predictive analysis (LPC) is used as the feature quantity for speech recognition.
ve coding) cepstrum or cepstrum. Cepstrum c is defined by the following equation (1).

[0007]

[Equation 1] c shown in equation (1) is a feature used as an input vector in a recognition system based on normal template matching, and in SS methods etc., noise components are subtracted in advance through preprocessing. This is the estimated value of the cepstrum. At this time, the audio signal s is inversely estimated from the cepstrum c. Here, it is assumed that minute noise Δn is newly mixed into the audio signal s. This indicates that the noise power added to the voice fluctuates or that the power fluctuates due to noise power estimation error. Due to noise power fluctuations, s in equation (1) is replaced by s+Δn, and is expanded as follows.

[0008]

[Formula 2] Here, if we let the cepstrum change due to minute noise Δn be Δc, and further assume that F[Δn]≪F[s], then

0
009]

The Δc vector calculated from equation (2) is used to transform the input vector as the direction of change in the cepstrum when noise is added. The transformation of the input vector is
This is done by newly using a discrimination measure that considers Δc in the direction of change of the cepstrum. That is, when the conventional distance measure is d and the distance considering the influence of noise addition on the feature parameters is dn, the new distance measure is (3).
This is achieved by setting (distance)=d+α×dn. However, α is a weighting coefficient. Further, dn is a distance weighted in the direction of change of Δn, and is expressed by the following equation. dn=((r・Δc)2/σ2)+r2−(r・Δc)
2 (i.e.r=x-c) x: input vector c: reference vector Δc: normalized Δc σ: weight (>1) As shown in FIG.
The input vector (x mark) seen from the reference vector is decomposed into a directional component dh where the input vector is subject to fluctuations due to noise, and a component dv perpendicular to it, and takes a small value for changes in the Δc direction, and It is weighted so that it takes a large value in the direction of , and is transformed as shown by the ellipse. Note that in FIG. 3, arrows indicate the direction of change in the input vector due to fluctuations in noise power. By using such a distance measure, it is possible to obtain the effect of adding a modification to the input vector in consideration of noise power fluctuations, and it is possible to improve recognition performance. That is, as shown in FIG. 4, the input vector 41 after transformation has redundancy compared to the input vector obtained by the conventional method (the part drawn in a circle with a dotted line), and the direction of change of the feature vector due to the change in noise power is (direction of arrow).

Next, the configuration and functions of the speech recognition device of this embodiment will be described. FIG. 1 is an explanatory diagram showing the processing flow of a speech recognition method in noise according to an embodiment of the present invention, and FIG. 5 is a block diagram of a speech recognition apparatus according to an embodiment of the present invention. In FIG. 5, 1 is an input device for inputting audio, 2 is a processing device such as a CPU, 3 is an output device for outputting the identification result of the input audio, and 4 stores standard patterns used for matching processing. It is an external storage device. The processing device 2 also includes a preprocessing unit 2a that performs preprocessing such as LPC spectrum analysis on input audio, and a feature extraction unit 2 that extracts feature vectors from the preprocessed input audio.
b, and an identification unit 2c that transforms the input vector in consideration of background noise and identifies the input speech based on the distance between the modified input vector and the feature vector of the standard pattern. The feature vectors of the standard patterns registered in the external storage device 4 are extracted from input speech (speech+noise) using the multi-template method. The identification unit 2c also detects minute noise power (background noise power) in the input voice.
The direction of movement in the case where . With such a configuration, the processing shown in FIG. 1 is performed. That is, the preprocessing unit 2a performs LPC spectrum analysis on the input speech (speech + noise) (101
), the feature vector is extracted by the feature extraction unit 2b (1
02). Furthermore, the distance d shown in the above equation (3) is determined by the identification unit 2c (103). On the other hand, the identification unit 2c calculates the direction of change of the input vector in consideration of minute fluctuations in the background noise power (104). In this process, the direction of change Δc of the cepstrum when minute background noise is added to the input speech is determined (104a), and the distance dn is determined in consideration of the influence of the noise power on the feature parameters (104b). Furthermore, from the distance d calculated assuming speech + noise and the distance dn calculated taking into account minute fluctuations in background noise power, a new distance measure that takes into account the changing direction Δc of the cepstrum is calculated, and according to this, the input vector is transformed and subjected to matching processing with the feature vector of the standard pattern (105), and the identification result is output (106). According to this embodiment, since the input vector is directly transformed in consideration of minute fluctuations in background noise power, the processing time can be reduced compared to, for example, a method of transforming a preset standard pattern.

[0011]

[Effects of the Invention] According to the present invention, in a speech recognition method based on pattern matching, it is possible to reduce the reduction in the identification rate caused by estimation errors in noise power caused by temporal changes in background noise. , it is possible to suppress a large increase in processing time. For example, in the multi-template method, the SN ratio of the input audio changes to some extent, so
Complement between templates becomes possible, and input samples are less likely to be erroneously recognized in an identification category with an SN ratio different from that of the input voice. Furthermore, if the identification rate is not changed, the number of templates can be reduced. Furthermore, with the SS method, it is possible to reduce excess or deficiency in the amount of noise removed due to noise power estimation error.

0012

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing the processing flow of a speech recognition method in noise according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing the processing flow of a conventional speech recognition method in noise.

FIG. 3 is an explanatory diagram showing weighting of input vectors in a speech recognition method under noise according to an embodiment of the present invention.

FIG. 4 is an explanatory diagram showing the direction of change of an input vector in a speech recognition method under noise according to an embodiment of the present invention.

FIG. 5 is a configuration diagram of a speech recognition device in an embodiment of the present invention.

[Explanation of symbols]

1 Input device 2 Processing equipment 2a Pre-processing section 2b Feature extraction section 2c Identification part 3 Output device 4 External storage device 41 Input vector after transformation

Claims (1)

[Claims] Claim 1: A speech recognition method for a system having a recognition system based on pattern matching, which uses spectral information of noise mixed in input speech to calculate the direction in which an input vector changes due to minute fluctuations in noise power. , setting a weighted distance measure so that the distance in the direction of change is smaller than the direction perpendicular to the direction of change, transforming the input vector in the direction of change in noise power, and identifying speech by the transformed input vector. A speech recognition method in noise characterized by: