JPH01274198A - Speech recognition device - Google Patents

Abstract

PURPOSE:To improve the recognition rate in the presence of a large noise by selecting the best noiseless feature vector time series among noiseless feature vector candidates based upon the minimization of distortion with a reference pattern at the time of pattern matching. CONSTITUTION:A vector quantizing means (ambiguity vector quantizer 20) inputs a speech signal on which a noise is superposed and performs the vector quantization of a feature vector time series according to a noise added code table 6, and a reverse vector quantizing means (reverse ambiguity vector quantizer 22) performs the reverse vector quantization of label candidate time series according to a noiseless code table 9. Then a multiple vector recognition processing means (multiple vector recognition processing circuit 26) select a noiseless feature vector time series among input noiseless feature vector candidate time series and performs speech recognition processing. Consequently, the feature vector of a speech is not distorted by the noise removal processing and the recognition rate of the speech does not decrease even in the presence of a loud noise.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device that recognizes input speech and has a noise removal function that suppresses noise superimposed on an input speech signal.

[Conventional technology]

In speech recognition using speech spectral 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. Signal processing techniques that attempt to improve the S/N ratio by removing only noise from a voice signal that already contains noise are called noise suppression, noise removal, voice enhancement, etc., and many methods have been proposed. There is. Recently, as a noise suppression method based on a new concept, the 5pectral mapping method removes noise from the standpoint of signal detection using a known mapping relationship between a noise-superimposed signal space and a noise-free signal space generated using vector quantization. The method called Biing-Hwang Juan
g+L, RoRabiner, SignalRes
toration by 5pectral Mapp
ing method, 1987 IEEE INTERNATION
AL C0NFERENCE ON ACUSTIC
3,5PEECH.

& 5IGNAL PROCESSING Volume
e 4 PP, 6.6.1-6.6.4 April 1
987. Dallas J (hereinafter this paper will be referred to as Document [1]), and is considered to be effective as a noise suppression method for voice transmission.

As a configuration example of a speech recognition device incorporating a noise removal circuit using this 5pectral mapping method, the second
A diagram can be considered. There are various recognition methods in recognition devices, but they have word-based templates and DP (Dynam
ic Programming) A discrete word recognition device using matching will be explained as an example.

In FIG. 2, 1 is an input terminal for audio signals, 2 is an input audio signal, 3 is an analysis circuit for acoustically analyzing the input audio signal 2, and 4
is a feature vector time series, 5 is a vector quantizer that vector-quantizes the feature vector time series output from the analysis circuit using a noise-added codebook 6, 7 is a label candidate time series of codewords, and 8 is a vector quantizer An inverse vector quantizer that performs inverse vector quantization on the label candidate time series 7, which is the output of 5, using a noise-free codebook 9; 10 is a noise-free feature vector time series; 11 is shown in the above-mentioned document [1]. This 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 noise-free feature vector time series 10 and the reference button 15 expressed by the noise-free feature vector time series output from the template memory 14;
Bump matching circuit that outputs matching distortion 16, 1
3 is a recognition control circuit that specifies a reference pattern in baton matching and outputs a recognition result; 17 is a recognition control circuit 1;
3 to specify the reference pattern
4 is the address data to be sent, 18 is the recognition result, and 19 is a recognition processing circuit composed of a slam 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 to make the noise mode the same as that of the voice signal, for example, adding noise in the time waveform domain so that the signal-to-noise ratio is the same as the noise superimposed human voice signal 2, and reanalyzing this to add noise. It consists of code words generated by converting them into feature vectors. The input audio signal 2 inputted to the input terminal 1 and superimposed with noise is acoustically analyzed by the analysis circuit 3, and a feature vector time series 4 is obtained (X(nl j
n = 1, 2. -, N).

Here, N indicates the number of feature vectors. Noise removal circuit 11
In the vector quantizer 5, the feature vector time series 4
As input, X (n
) with all the codewords in the noise-added codebook 6, and the labels of the codewords from the one with the highest likelihood to the highest rank are used as label candidates (m4(n)l i = 1.2.-, L), this is calculated for n = 1.2°-, N, and the label candidate time series 7 is (M(nl: n = 1, 2
．． --, N) (where M (n) = (m=
(n) l i = 1. 2. --, l)). L is an integer of 1 or 2 or more. The inverse vector quantizer 8 is the label candidate time series 7 (M(
nl i n = 1, 2. --, N) (
However, M(n) = (mi(nl l + = 1,
2. ---, L) ), label candidate M(nl= (m, (nl l i = 1.2
．．・-, L) is inverse vector quantized using the noiseless codebook 9 and a noiseless feature vector candidate (Z; (nl l i
= 1, 2, -.

L), and calculate these L noise-free feature vector candidates (Z
+ (nl l i =1.2.--, find Y (n) as the average vector of Ll.

That is, L yt(nl= □ Σ2, (nl (Equation 1)%
Formula % Here, yt(n), ztl, (n) are respectively Y
(n), Z, (nl-th dimension component.

In the recognition processing circuit 19, the template memory 14 is a reference pattern 15 specified by the address data 17 output from the recognition control circuit 13 (T(kllk=1.2.
-..., Kl (K is the number of feature vectors) are sent to the matching circuit 12. Battan matching circuit 12
is the above reference button 15 (T(k)l k=1.2
．． -1K) and the noise-free feature vector time series 10 which is the output of the noise removal circuit 11 (Y(nl l
DP matching is performed with n = 1, 2, -, N). For example, the recurrence formula for DP matching is as follows.

g (k, n-1) + D (k, n) g (k, n) -min g (k-1, n-1) +
2XD(k,n)g(k-1,n)+D(k,n) (Formula 2) Here, D(k,n) is the distortion between the feature vector T(kl and the feature vector Y(n)) (2) takes the case of DP matching without slope restriction as an example. g(K, N
) is normalized by dividing by the urbanization distance, and the matching distortion 1
Output as 6. The recognition control circuit 13 sequentially changes the reference patterns specified by the address data 17 and selects the label of the reference pattern that minimizes the matching distortion from the matching distortion 16 output by the button matching circuit 12 for each reference pattern as the recognition result 18. Output.

[Problem to be solved by the invention]

The mapping relationship between the noise-superimposed signal space and the noise-free signal space, which is the basis of noise removal using the 5-pectral mapping method, can be achieved by creating the noise-added codebook 6 by adding processing equivalent to noise superimposition to the noise-free codebook 9. It is formed. However, since actual noise has statistical variance, errors occur in the mapping relationship due to the difference between the noise equivalently added in generating the mapping relationship and the noise superimposed on the human voice signal 2. Furthermore, as the level of noise superimposed on the input audio signal 2 increases, the spectral envelope of the input audio signal 2 becomes smoother. As a result, the phonological characteristics included in the feature vector time series 4 output from the analysis circuit 3 disappear, and the above-mentioned mapping error increases significantly. For example, when the number of label candidates in the noise removal circuit 11 is set to 1, if a mapping error occurs due to noise dispersion, the distortion due to the mapping error will be the matching distortion 16 in the slam matching.
Therefore, as the level of superimposed noise increases, recognition performance decreases rapidly. Furthermore, if you increase the number of candidates by setting L to a number greater than or equal to 2, the probability that a candidate based on the correct mapping will be included among the multiple candidates will increase, but which of the multiple candidates is the candidate based on the correct mapping? I don't know.

Therefore, the above-mentioned document [1] uses an improvement measure in which the average of feature vectors of a plurality of candidates is output. However, due to this averaging process, components of the incorrect candidate feature vectors are mixed into the candidate feature vectors that should have been selected, causing distortion and deteriorating recognition performance. Therefore, the above document [1
] method is effective in improving the auditory signal-to-noise ratio in speech transmission, but is ineffective when applied to speech recognition. As described above, the conventional speech recognition apparatus has a problem in that the noise removal process causes distortion in the speech feature vector, and the speech recognition rate decreases under high noise.

The present invention has been made in order to solve the above-mentioned problems, and provides a speech recognition device that does not distort the speech feature vector through noise removal processing and does not reduce the speech recognition rate even under high noise. The purpose is to

[Means to solve the problem]

The speech recognition device according to the present invention includes a noise-free codebook 9 whose codewords are feature vectors of speech on which no noise is superimposed, and a process equivalent to noise superimposition on each codeword of the noise-free codebook 9. A vector quantizer that vector quantizes the generated noise-added codebook 6 and the feature vector time series of the input speech signal on which noise is superimposed according to the noise-added codebook 6 and outputs a plurality of label candidate time series indicating labels of code words. quantization means (ambiguous vector quantizer 20) and a plurality of label candidate time series, which are output signals of this vector quantization means (ambiguous vector quantizer 20), are subjected to inverse vector quantization according to the noiseless codebook 9, and a plurality of label candidate time series are An inverse vector quantization means (inverse ambiguous vector quantizer 22) that outputs a noise-free feature vector candidate time series, and a plurality of noise-free signals that are output signals of this inverse vector quantization means (inverse ambiguous vector quantizer 22) The present invention is characterized by comprising a multi-vector recognition processing means (multi-vector recognition processing circuit 26) that inputs a feature vector candidate time series, selects a noise-free feature vector time series, and performs speech recognition processing.

[Effect]

The vector quantization means (ambiguous vector quantizer 20) is
A speech signal on which noise is superimposed is manually processed, the feature vector time series of this input speech signal is vector quantized based on the noise-added codebook 6, and a plurality of label candidate time series are subjected to inverse vector quantization means (inverse ambiguous vector quantization). 22). The inverse vector quantization means (inverse ambiguous vector quantizer 22) performs inverse vector quantization on the input plurality of label candidate time series based on the noiseless codebook 9, and multiplexes the plurality of noiseless feature vector candidate time series. The vector recognition processing means (multiple vector recognition processing circuit 26) is provided with the vector recognition processing means (multiple vector recognition processing circuit 26). The multiple vector recognition processing means (multiple vector recognition processing circuit 26) selects a noise-free feature vector time series from a plurality of input noise-free feature vector candidate time series and performs speech recognition processing.

[Embodiments of the invention]

FIG. 1 is a block diagram showing the configuration of a speech recognition device according to an embodiment of the present invention. In FIG. 1, components corresponding to those shown in FIG. 2 are given the same reference numerals, and their explanations will be omitted. This embodiment will be explained by taking as an example an M-word speech recognition device that performs recognition by DP matching with a word-by-word template, as in the conventional example.

In FIG. 1, 20 is a vector quantization means that vector quantizes the feature vector time series 4, which is the output signal of the analysis circuit 3, using a noise-added codebook 9, and outputs a plurality of label candidate time series 21. The ambiguous vector quantizer 22 is an inverse vector quantizer that performs inverse vector quantization on a plurality of label candidate time series 21 using a noiseless codebook 9 and outputs a plurality of noiseless feature hector candidate time series 23. Vector quantizer, 24 is a noise-added codebook 6
A noise-free multiple vector generation circuit consisting of a noise-free codebook 9, an ambiguous vector quantizer 20, and an inverse ambiguous vector quantizer 22; 25 is a reference button 15 using a plurality of noise-free feature vector candidate time series 23 as input; 26 is a multi-vector bump matching circuit 25, a template memory 14, and a recognition control circuit 1.
3. This is a multi-vector recognition processing circuit as a multi-vector recognition processing means consisting of 3 and 3.

This speech recognition device uses fuzzy vector quantization (Fuzzy
Vector Quantization)
Noise-free multiplexing employs a new method that simultaneously performs input signal noise removal and button recognition, and outputs a plurality of noise-free feature vector candidate time series 23 by inputting the feature vectors of the input audio signal with superimposed noise. It includes a vector generation circuit 24 and a multiple vector recognition processing circuit 26 that performs recognition by inputting a plurality of noise-free feature vector candidate time series 23, which are output signals of the noise-free multiple vector generation circuit 24. As described above, the noise-free multi-vector generation circuit 24 generates a noise-free codebook 9 whose codewords are voice feature vectors on which no noise is superimposed, and a noise-free multivector generator 24 that generates a noise-free codebook 9 in which codewords are feature vectors of speech on which no noise is superimposed, and a noise-free multivector generating circuit 24 that generates a noise-free codebook 9 in which codewords are feature vectors of speech on which no noise is superimposed. An ambiguous method that vector quantizes the noise-added codebook 6 generated by performing processing equivalent to the noise-added codebook 6 and the feature vector time series 4 of the input speech signal on which noise is superimposed according to the noise-added codebook 6 to output a plurality of label candidate time series 21. Vector quantizer 20 and ambiguous vector quantizer 2o
A plurality of label candidate time series 21 which are the output signals of
and an inverse ambiguous vector quantizer 22 that performs inverse vector quantization on the noise-free feature vector candidate time series 23 according to the noise-free codebook 9 and outputs a plurality of noise-free feature vector candidate time series 23.

The multiple vector recognition processing circuit 26 processes a plurality of noise-free feature vector candidate time series (Z(nNn=1,2°·-,N) (where Z(n) = (Z; fn) I i = 1,
2゜-・-2L)), and the noise-free feature vector candidate Z(n)=(Zt fn) l at arbitrary n.
i = 1. 2°-, L), by selecting the candidate that maximizes the likelihood with the reference pattern for each reference pattern as the optimal noise-free feature vector.
The reference button that finally gives the maximum likelihood for all time series n-N is determined to be the recognition category, and at the same time, processing is performed to determine the optimal noise-free feature vector time series.

The operation of this embodiment will be explained.

The ambiguous vector quantizer 20 is a feature vector time series 4 of the input speech (X(nl l n= 1.2.-.

N) as input, and the feature vector X (
Find the precedence between nl and all codewords in the noise-added codebook, and select the labels of the codewords from the one with the highest likelihood to the highest likelihood as label candidates (mt (n) l i = 1. 2. --- ,
L), and n=1.2.・−, N, multiple label candidate time series 21 (M(nl l n=
1. 2. −.

Nl (where M(n)=(ml(n)11=1.2
．． -・-9L)). The inverse ambiguity vector quantizer 22 generates a plurality of label candidate time series 21 (M(
nl i n = 1, 2. ---, Nl (
However, M (n) = (rn t(nl l i =
1, 2. --, L) ) label candidate M(n) = (mt (n) l i =
1, 2, -.

L) is inverse vector quantized using the noiseless codebook 9, and multiple noiseless feature vector candidates (Zt (n) l i =
1.2. -..., L) and convert this to n= 1.2. −
--.

N, and a plurality of noise-free feature vector candidate time series 23 (Z(n) l n = 1, 2,
--, Nl (where z(n)= (Z! (nl I
i = 1. 2. -, L)).

The multi-vector bump matching circuit 25 generates a plurality of noise-free feature vector candidate time series 23, which are the output signals of the inverse ambiguous vector quantizer 22 (Z(n) l
n = 1, 2, -, N) (however, Z (n
) = (Zz (nl l i = 1. 2.--,
), and input reference pattern 15 (T(kl
lk=1. 2. .--, K) (K is the number of feature vectors) DP matching (Equation 2) is performed. However, since the multiple noise-free feature vector candidate time series 23 has L feature vector candidates for any n, the distortion D(k, n) at any k and n is defined as follows.

D(k, n) = min (d(T (kl, Z
t (nl))1≦i5L (Formula 3) where d(*, *) represents the distortion between the feature vectors. Thereby, the optimal feature vector for the reference pattern 15 is selected from among the plurality of candidates. As in the conventional example, g (K, N) obtained in the recurrence formula (Formula 2) is normalized by dividing it by the urbanization distance and output as matching distortion 16. According to the principle of DP matching, if the partial distortion of (Equation 3) is minimized, the matching distortion of the entire batten is minimized. The recognition control circuit 13 sequentially changes the reference pattern in the template memory 14 specified by the address data 17,
The matching distortion 16 output by the baton matching circuit 25 is determined for each reference pattern, and the label of the reference pattern that minimizes the matching distortion 16 is output as the recognition result 18. With respect to the reference pattern that minimizes the matching distortion 16, the vector time series optimally selected from among the multiple noise-free feature vector candidate time series 23 is a speech signal from which noise has been correctly removed from the input speech with superimposed noise. It approximates the noise-free feature vector time series, and the minimum matching distortion corresponds to the matching distortion between the noise-free input speech and the noise-free self-category template.

Note that the DP matching according to Equation 2 is given as an example, and does not limit the scope of application of the present invention.

In this way, in the noise-free multiple vector generation circuit 24,
The feature vector time series 4 of the input speech on which noise is superimposed is ambiguously mapped onto a plurality of noise-free feature vector candidate time series 23 containing correct noise-removed signals with redundancy, and the multiple vector recognition processing circuit 26 processes the noise-free feature vector time series 4 . Recognition processing is performed by automatically selecting a noise-free feature vector time series that minimizes the amount of mapping by maximizing the likelihood using the reference pattern.

Therefore, this embodiment eliminates distortion caused by mapping errors and distortion caused by averaging feature vectors of multiple candidates, which are the problems of the conventional device described above, and the recognition rate does not decrease due to these. .

Above are the word-by-word templates. The present invention has been explained using a discrete word speech recognition device as an example, but the recognition units are not limited to words, but include CV (consonant-vowel), ■C■ (vowel-consonant-vowel), CVC (consonant-vowel-consonant), single syllable, morpheme, or other arbitrary units may be used, and as long as the template is registered in time series of feature vectors, the same effect as in the above embodiment can be achieved. In addition, HM M other than the recognition method using template matching
In statistical recognition methods such as (Ilidden Markov Model), °ζ also uses a transition probability model instead of a reference pattern, and for each transition probability model, a noise-free feature vector that maximizes the generation probability is used as a noise-free feature vector candidate. A similar effect can be obtained by selecting from.

Furthermore, for continuously uttered sounds other than discrete word sounds, a segmentation circuit that cuts out the continuous utterances into registered sound units is installed in the front stage of this With, and the cut out sounds can be used as the human power of this device. It is applicable. Furthermore, it goes without saying that the means for realizing the present invention is not limited to dedicated hardware, but can also be realized by software processing in a general-purpose computer or signal processing processor. Furthermore, the present invention can be expanded and applied to recognition devices for acoustic signals other than voice, image signals, text/graphic signals, and the like.

〔Effect of the invention〕

As described above, according to the present invention, there is provided a vector quantization unit that vector quantizes a feature vector time series of an input audio signal on which noise is superimposed according to a noise-added codebook and outputs a plurality of label candidate time series, and a vector quantization unit that outputs a plurality of label candidate time series. an inverse vector quantization means for inverse vector quantizing a plurality of label candidate time series, which are output signals of the means, according to the noiseless codebook and outputting a plurality of noiseless feature vector candidate time series; and an output of the inverse vector quantization means. The multi-vector recognition processing means inputs a plurality of noise-free feature vector candidate time series as signals, selects the noise-free feature vector time series, and performs speech recognition processing, so it can be referred to during slam munching. The optimal noise-free feature vector time series can be selected from among multiple candidates for noise-free feature vectors based on minimizing the distortion with the noise removal process. Therefore, even under high noise conditions, the recognition rate is extremely high.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the structure of a speech recognition device according to an embodiment of the present invention, and FIG. 2 is a block diagram showing the structure of a conventional speech recognition device. 6...Noise-added codebook, 9...Noise-free codebook, 2
0... Ambiguous vector quantizer (vector quantization means)
, 22... Inverse ambiguous vector generator (inverse vector quantization means), 26... Multiple vector recognition processing circuit (multiple vector recognition processing means). Agent Masu Oiwa A (and 2 others) Procedural amendment (voluntary) May 2011. 9th, 3rd, Representative of the person making the amendment: Moriya Shiki 5 Detailed description of claim 1 invention subject to amendment. 6. Contents of amendment (1) The scope of claims will be amended as shown in the attached sheet. (2) In the 3rd line of page 4 of the specification, replace ``chronological order'' with ``
"Time series 4". (3) In the 5th line of page 4 of the same book, the phrase ``code word'' is replaced with ``
It is corrected as "consisting of labels representing code words." (4) 2r on page 7, lines 14 to 16 of the same book.
g(k,n-1)+D(k,n)g(k
, n)=min g(k-1, n-1)+2XD(
k, n) g (k-1, n) + D (k, n) J
I'll correct that. (4) "Urbanization distance" on page 8, line 2 of the same book and page 17, line 9 is corrected to "urbanization distance (+N) J." (5) Same book, page 9, "urbanization distance" In the 20th line, replace "conventional" with "
It has a conventional noise removal function.'' (6) In the same book, page 10, line 16, the phrase ``indicates the label of the code word'' is corrected to ``consists of the label representing the code word.'' (7) On page 11, lines 12 and 13 of the same book, the phrase "input an audio signal superimposed with noise, and correct this input audio signal" is replaced with "J of the input audio signal superimposed with noise". (8) In the same book, page 12, line 16, "Codebook 9" is corrected to "Codebook 6." (9) In the same book, page 13, line 7 to line 8, the phrase "matching distortion" is corrected to "matching distortion 16." (10) In the same book, page 14, line 8, "input audio signal" is corrected to read "r input audio signal 2". (11) In the same book, page 14, line 19, the phrase "time series" is corrected to "time series 23." (12) “Noise-added codebook” on page 15, line 14 of the same book.
The text has been corrected to read "noise-added codebook 6." J裟YO 2, Claims In a speech recognition device that recognizes human speech and has a noise removal function that suppresses noise superimposed on an input speech signal, a feature vector of speech on which no noise is superimposed is used as a code word. A noise-free codebook, a noise-added codebook generated by performing processing equivalent to noise superimposition on each code word of this noise-free codebook, and a noise-added codebook that generates the feature vector time series of the noise-superimposed input speech signal. A vector quantization means that outputs a plurality of label candidate time series, and a plurality of label candidates that are output signals of this vector quantization means. Inverse vector quantization means for inverse vector quantization of a time series according to the noiseless codebook and outputting a plurality of noiseless feature vector candidate time series, and a plurality of noiseless feature vector candidates which are output signals of the inverse vector quantization means. A speech recognition device comprising: multi-vector recognition processing means for inputting a time series, selecting a noise-free feature vector time series, and performing speech recognition processing.

Claims (1)

[Claims] In a speech recognition device that recognizes input speech and has a noise removal function that suppresses noise superimposed on an input speech signal, a noise-free codebook that uses feature vectors of speech on which no noise is superimposed as codewords, and this noise A noise-added codebook is generated by applying processing equivalent to noise superimposition to each codeword of the non-codebook, and the feature vector time series of the input speech signal with noise superimposed is vector quantized according to the above-mentioned noise-added codebook, and the codewords are A vector quantization means for outputting a plurality of label candidate time series indicating labels, and a plurality of label candidate time series, which are output signals of the vector quantization means, are inverse vector quantized according to the noiseless codebook to generate a plurality of noiseless features. An inverse vector quantization means that outputs a vector candidate time series, and a plurality of noise-free feature vector candidate time series that are output signals of this inverse vector quantization means are input, and a noise-free feature vector time series is selected to perform speech recognition processing. 1. A speech recognition device comprising: multi-vector recognition processing means for performing.