JPH04298797A - Voice recognition device - Google Patents

Abstract

PURPOSE:To obtain the voice recognition device which performs a noise removing process without averaging the feature vector itself of a voice and does not decrease in the recognition ratio of the voice even in noisy environment as to a voice recognition device which has a noise removing function for suppressing a noise superposed on an input voice signal and recognizes the input voice. CONSTITUTION:A proximate code word selector 20 outputs plural proximate code word candidates 21 which have short distance to the input voice signal and distance candidates 22 between the respective proximate code word candidates and input voice signal from a noise addition code book 6. A moving vector calculator 23 calculates moving vector candidates 24 of the proximate code word candidates before and after noise superposition. A noise removal unit 25 inputs the moving vector candidates 24, distance candidates, and the feature vector 4 of the input voice signal and corrects the movement of the feature vector of the input voice signal caused by the noise superposition by using the load mean of the moving vector candidates to remove the noise.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise-resistant speech recognition device having a noise removal function for suppressing noise superimposed on an input speech signal.

0002

2. Description of the Related Art In speech recognition using speech spectrum information, deformation of the speech spectrum due to noise superposition significantly reduces recognition performance. Therefore, improving the resistance to noise is an important issue in putting speech recognition devices into practical use. Noise-cancelling microphones are often used to suppress noise contamination caused by environmental noise, but there are still cases where a sufficient signal-to-noise ratio (S/N) cannot be obtained or noise is superimposed during the audio signal transmission process. However, the signal processing technology that attempts to improve the S/N by removing only the noise from the audio signal that already contains noise is known as noise suppression/
Many methods have been proposed, called noise removal and speech enhancement.

[0003]Recently, as a noise suppression method based on a new concept, a method called the spectral mapping method is used to remove noise using a known mapping relationship between a noise-superimposed signal space and a noise-free signal space generated using vector quantization. published the paper “Biing-Hwang Jang, L.
．． R. Rabiner, ■Signal Restru
cation by Spectral Mapping
■, 1987, IEEE INTERNATION
AL CONFERENCE ON ACUSTIC
S, SPEECH AND SIGNAL PROCE
SSINGS. , Volume 4, PP. 6.6
．． 1-6.6.4, April 1987, Dal
(Hereinafter, this paper will be cited as Reference [1])
This method was proposed in 2013, and is said to be effective as a noise suppression method for voice transmission.

FIG. 2 shows an example of the configuration of a speech recognition device incorporating a noise removal circuit using the spectral mapping method. There are various recognition methods in recognition devices, but they have word-based templates and DP (Dynami
An example of an isolated word recognition device using matching will be described.

In FIG. 2, 1 is an audio signal input terminal;
3 is an input audio signal, 3 is an analysis circuit that acoustically analyzes the input audio signal 2, 4 is a feature vector, and 5 is a vector quantizer that vector-quantizes the feature vector 4 output from the analysis circuit using a noise-added codebook 6. 7 is a label candidate consisting of a label representing a code word; 8 is an inverse vector quantizer that performs inverse vector quantization of the label candidate 7, which is the output of the vector quantizer 5, using a noiseless codebook 9; and 10 is an inverse vector quantizer The noise-free feature vector 11 is a noise removal circuit composed of a vector quantizer 5, a noise-added codebook 6, an inverse vector quantizer 8, and a noise-free codebook 9. 12 performs DP matching between the time series of the noise-free feature vectors 10 and the reference pattern 15 expressed as the time series of the noise-free feature vectors output from the template memory 14, and creates a matching distortion 16.
13 is a recognition control circuit that specifies a reference pattern in pattern matching and outputs a recognition result; 17 is address data that the recognition control circuit 13 sends to the template memory 14 to specify a reference pattern; 18 is a recognition result, and 19 is a recognition processing circuit composed of a pattern matching circuit 12, a template memory 14, and a recognition control circuit 13.

Next, the operation will be explained. In the noise removal circuit 11, the noise-free codebook 9 is configured with speech feature vectors on which no noise is superimposed as codewords, and the noise-added codebook 6 is configured with noise-superimposed input for each codeword of the noise-free codebook 9. Processing that makes the speech signal noise-like, for example, adding noise in the time waveform domain so that the signal-to-noise ratio is the same as the noise superimposed input speech signal 2, and reanalyzing this to create a noise-added feature vector. It consists of a codebook generated by converting to . The input audio signal 2 with superimposed noise input to the input terminal 1 is acoustically analyzed by the analysis circuit 3, and the feature vector 4 is X(k) (k represents a time series number. k=
1, 2, . ．． ．． , K).

In the noise removal circuit 11, the vector quantizer 5 inputs the feature vector 4, X(k), calculates the distances to all the code words in the noise-added codebook 6, and selects the M-th one from the one with the smallest distance. Label candidate Ln(k
) = {lni(k) | i = 1, 2, . ．． ．． ,M
}. M is an integer of 1 or 2 or more.

The inverse vector quantizer 8 calculates the label candidate 7 Ln(k) = {lni(k) | i = 1,2
、． ．． ．． , M} is inverse vector quantized using the noise-free codebook 9 to create a feature vector candidate {vci(k) | i =1,2
、． ．． ．． , M}, and calculate these M feature vector candidates {
vci(k) |i = 1, 2, . ．． ．． , M}, a noise-free feature vector Y(k) is obtained. That is,

[0009]

[Math 1]

In the recognition processing circuit 19, the template memory 14 stores address data 1 output from the recognition control 13.
The reference pattern 15 specified by 7 is {T(l) |
l = 1, 2, . ．． ．． , L} (L is the number of feature vectors) to the pattern matching circuit 12. The pattern matching circuit 12 uses the above reference pattern 15 {T(l)
| l = 1, 2, . ．． ．． , L} and noise removal circuit 1
{Y(k) | k = 1, 2, . ．． ．． ，
Perform DP matching with K}. For example, the recurrence formula for DP matching is as follows.

[0011]

[Math 2]

Here, g(k,l) is the feature vector Y(
k) and the feature vector T(l). The above equation takes as an example the case of DP matching without slope restriction. In this recurrence formula (Equation 2), G(K,L) is the reference pattern 15 described above, {T(l) | l= 1
,2,. ．． ．． , L} and {Y(
k) | k = 1, 2, . ．． ．． , K}, which is the sum of the time series lengths (K+L), is normalized and output as matching distortion 16.

The recognition control circuit 13 receives address data 17
The matching distortion 1 outputted by the pattern matching circuit 12 for each reference pattern by sequentially changing the reference pattern specified by
6, the label of the reference pattern that minimizes the matching distortion is output as the recognition result 18.

[0014]

[Problem to be solved by the invention] Spectral M
The mapping relationship between the noise-superimposed signal space and the noise-free signal space, which is the basis of noise removal using the appping method, is formed by creating the noise-added codebook 6 by adding processing equivalent to noise-superimposition to the noise-free codebook 9. Ru. However, as the level of noise superimposed on the input audio signal 2 increases, the spectral envelope of the input audio signal 2 flattens. Therefore, the phonological characteristics included in the feature vector time series 4 output from the analysis circuit 3 are reduced, and when vector quantization is performed using the noise-added codebook 6, there is a possibility that the vector quantization will be performed into a code word by correct mapping. The mapping error increases significantly. For example, when the number M of label candidates in the noise removal circuit 11 is set to 1, when a mapping error occurs, the distortion due to the mapping error is directly reflected in the matching distortion 16 in pattern matching, so as the level of superimposed noise increases, the recognition performance increases. decreases rapidly. Therefore, the prior art uses an improvement measure to increase the number of candidates by setting M to a number greater than or equal to 2. This increases the probability that a correct mapping candidate will be included among the plurality of candidates, but it is not known which one of the plurality of candidates is the correct mapping candidate. Therefore, a method is used in which the average of feature vectors of multiple candidates is output. However, due to this averaging process,
Because components of the feature vector of the incorrect candidate are mixed into the feature vector of the candidate that should have been selected, the phonological distinctiveness is blurred.
Recognition performance deteriorates. Therefore, although the conventional technology is effective in improving the auditory signal-to-noise ratio in speech transmission, it is less effective when applied to speech recognition. As described above, conventional speech recognition devices with a noise removal function have the problem that the speech feature vectors are averaged during the noise removal process, which blurs the phonological features and reduces the speech recognition rate under high noise conditions. was there.

The present invention has been made to solve the above-mentioned problems, and provides speech recognition in which the speech feature vectors are not averaged through noise removal processing, and the speech recognition rate does not decrease even under high noise. The purpose is to obtain equipment.

[0016]

[Means for Solving the Problems] A speech recognition device according to the present invention includes a noise-free codebook whose codewords are feature vectors of speech on which no noise is superimposed, and a noise-free codebook in which each codeword of the noise-free codebook has a noise-free codebook. A noise-added codebook generated by processing equivalent to superposition, a plurality of neighboring codeword candidates from this noise-added codebook whose distance to the input speech signal is small, and the relationship between each neighboring codeword candidate and the input speech signal. a nearby code word selection means for outputting a distance candidate of the above-mentioned neighborhood code word; a movement vector calculation means for calculating a movement vector of the neighboring code word before and after noise superimposition; The motion vector candidate, the distance candidate which is the output signal of the neighboring code word selection means, and the feature vector of the noise-superimposed input speech signal are input, and the weighted average of the motion vector candidate is used to calculate the input speech resulting from the noise superposition. It is equipped with means for correcting the movement of the signal feature vector and removing noise.

[0017]

[Operation] The noise removal means of the present invention uses a weighted average of a plurality of movement vector candidates, which are output signals of the movement vector calculation means, to correct movement of the feature vector of the input speech signal caused by noise superposition. Since noise is removed, the noise can be removed without averaging the feature vectors themselves.

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram showing the configuration of a speech recognition apparatus according to an embodiment of the present invention. In Figure 1, Figure 2
Components corresponding to those shown in are given the same reference numerals, and their explanations will be omitted. This embodiment, like the conventional example,
An example of an isolated word recognition device that performs recognition by DP matching with a word-by-word template will be described.

In FIG. 1, reference numeral 20 denotes a plurality of neighboring codeword candidates 21 having a small distance from the feature vector 4 of the input speech signal from the noise-added codebook 6, and distance candidates between each neighboring codeword candidate and the input speech signal. This is a neighborhood code word selector that outputs 22. Here, the distance is, for example, the Euclidean distance to the feature vector 4 of the input audio signal. Reference numeral 23 denotes a movement vector calculator that calculates a movement vector candidate 24 of the neighboring codeword candidate 21, which is the output signal of the neighboring codeword selector 20, before and after noise superimposition. 25 is movement vector candidate 2
4 and the distance candidate 2 which is the output signal of the neighborhood code word selector 20
2 and a feature vector 4 which is the output signal of the analysis circuit 3 as inputs, and outputs a feature vector 27 from which noise has been removed. 26 is a noise-added codebook 6, a noise-free codebook 9, a neighboring codeword selector 20, and a movement vector calculator 23.
This is a noise removal circuit composed of a noise remover 25 and a noise remover 25.

The principle of the noise removal method of the speech recognition device according to the present invention is as follows. The noise-added codebook 6 processes each codeword of the noise-free codebook 9 so that it becomes the noise form of the noise-superimposed input speech signal, for example, so that it has the same signal-to-noise ratio as the noise-superimposed input speech signal 2. This can be obtained by adding noise in the temporal waveform domain, reanalyzing this, and converting it to a noise-added feature vector, but this process of noise superposition can be regarded as movement of each codeword in the feature space. can. Since the feature vector of each codeword is generally obtained by performing nonlinear processing from the time waveform domain, the movement due to noise superposition is different for each codeword. This movement in the feature space for each code word will be called a movement vector. In order to remove a noise component from a code word on which noise is superimposed, it is sufficient to correct the movement caused by the noise superposition, that is, to subtract the movement vector. Although the feature vector X(k) of the input speech does not completely match the code word of the noise-added codebook 6, the movement of the feature vector X(k) of the input speech caused by noise superposition is
can be considered to be close to the movement vector of the code word in the vicinity of . Therefore, in the speech recognition device according to the present invention, the feature vector X(k) of the input speech selected from the noise-added codebook 6 is
M codewords in the neighborhood of {vni(k) | i = 1
,2,. ．． ．． , M}, the noise is removed by subtracting the weighted average of the movement vectors.

Next, the operation of this embodiment will be explained. The neighborhood codeword selector 20 selects M neighborhood codeword candidates 21 from the noise-added codebook 6 that are close to x(k), which is the input speech signal 4, Vn(k) = {vni(k). )｜
i=1,2,. ．． ．． , M} and D(k)={d
i(k) | i = 1, 2, . ．． ．． , M} is output.

The movement vector calculator 23 uses the nearby code word candidate 21, which is the output signal of the nearby code word selector 20, that is, V
n(k)={vni(k)|i=1,2,. ．． ．． ，
M} and the code word of the noise-free codebook 9 are input, and B(k) = {bi(k)|i which is the movement vector candidate 24 before and after noise superimposition of the neighboring code word candidate 21 =
1, 2,. ．． ．． , M}. The movement vector candidate 24 is B(k)={bi(k)|i=1,
2,. ．． ．． , M} is calculated as follows.

[0023]

[Math 3]

Here, vci(k) is a codeword of the noiseless codebook 9, and vni(k) is a codeword of the noise-added codebook 6.

The noise remover 25 inputs a plurality of movement vector candidates 24, distance candidates 22 which are the output signals of the neighborhood code word selector 20, and feature vectors 4 which are the output signals of the analysis circuit 3, and calculates the noise-removed features. Z(k
) is output. Z(k), which is the noise-removed feature vector 27, is calculated as follows.

[0026]

[Math 4]

[0027]

[Math 5]

Here, wi is the movement vector candidate B(k)=
{bi(k)| i = 1, 2, . ．． ．． , M}, and p is a constant that determines the weight of the weighted average, for example, p=1.

Feature vector 27 subjected to the above noise removal
Z(k) is sent to the pattern matching circuit 12,
Enter recognition processing. Since the subsequent processing is completely equivalent to the method described in the section of the prior art, the explanation will be omitted.

As described above, in this embodiment, noise is removed from the viewpoint of movement of feature vectors, so that the phonological noise caused by averaging the feature vectors of multiple candidates, which is a problem with the conventional device described above, is eliminated. Blur is reduced, and a significant drop in recognition rate can be suppressed.

In this embodiment, a weighted average of the movement vectors is used, but in the present invention, all weight coefficients wi are set to 1.
That is, it includes a method using a simple average of moving vectors.

In this embodiment, the present invention has been explained by taking as an example an isolated word speech recognition device that registers word-based templates; however, the recognition unit is not limited to words;
It may be any unit such as a phoneme or a syllable. Further, although DP matching has been described as an example of a recognition method, any recognition method using a time series of feature vectors can be applied.

[0033] Furthermore, it can be used not only for speech recognition but also as a noise suppression method for speech transmission.

[0034]

As described above, according to the present invention, there is a noise-free codebook 9 whose codeword is a feature vector of speech on which no noise is superimposed, and a noise-free codebook 9 in which each codeword of this noise-free codebook 9 is superimposed with noise. A noise-added codebook 6 generated by performing processing equivalent to , a plurality of neighboring codeword candidates from the noise-added codebook 6 whose distance to the input speech signal is small, and a connection between each neighboring codeword candidate and the input speech signal. Neighborhood codeword selection means for outputting distance candidates; movement vector calculation means for calculating a movement vector of the neighborhood codeword before noise superimposition; and output signals of the movement vector calculation means for a plurality of movement vector candidates and neighbors. The distance candidate, which is the output signal of the code word selection means, and the feature vector time series of the noise-superimposed input speech signal are input, and the weighted average of the movement vector candidates is used to calculate the feature vector of the input speech signal resulting from the noise superposition. Since the system is equipped with a means to remove noise by correcting the movement of the image, the blurring of phonetic properties caused by averaging the feature vectors of multiple candidates is reduced, and a significant drop in recognition rate is suppressed even under high noise conditions. be able to.

[Brief explanation of drawings]

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

FIG. 2 is a block diagram showing the configuration of a conventional speech recognition device.

[Explanation of symbols]

1. Audio signal input terminal 2 Input audio signal 3 Analysis circuit 4 Feature vector 5 Vector quantizer 6 Noise-added codebook 7 Label candidates 8 Inverse vector quantizer 9 Noise-free codebook 10 Noise-free feature vector 11 Noise removal circuit 12 Pattern matching circuit 13 Recognition control circuit 14 Template memory 15 Reference pattern 16 Matching distortion 17 Address data 18 Recognition results 19 Recognition processing circuit 20 Neighborhood code word selector 21 Neighboring codeword candidates 22 Distance candidate 23 Movement vector calculator 24 Movement vector candidates 25 Noise remover 26 Noise removal circuit 27 Noise removed feature vector

Claims (1)

[Claims] Claim 1: A speech recognition device that recognizes input speech and has a noise removal function that suppresses noise superimposed on an input speech signal, wherein a noise-free code uses a feature vector of speech on which no noise is superimposed as a code word. a noise-added codebook generated by applying processing equivalent to noise superposition to each codeword of this noise-free codebook, and a plurality of neighboring codes from this noise-added codebook that have a small distance to the input audio signal. Neighborhood codeword selection means for outputting word candidates and distance candidates between each of the neighborhood codeword candidates and the input audio signal; movement vector calculation means for calculating a movement vector of the neighborhood codeword before noise superimposition; A plurality of movement vector candidates, which are the output signals of the vector calculation means, the distance candidates, which are the output signals of the neighborhood code word selection means, and the feature vector of the noise-superimposed input audio signal are input, and the weighted average of the movement vector candidates is used. 1. A speech recognition device comprising means for correcting movement of a feature vector of an input speech signal caused by noise superimposition and removing noise.