JPH0766271B2 - Voice recognizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice recognition device having a noise removal function of suppressing noise superimposed on an input voice signal and recognizing an input voice.

[Conventional technology]

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 for practical use of the voice recognition device. Noise-canceling microphones are often used to suppress noise mixing due to environmental noise, but there are cases in which a sufficient signal-to-noise ratio (S / N) cannot be obtained, or noise is superimposed during the audio signal transmission process. There is a signal processing technology that attempts to improve S / N by removing only noise from a speech signal that already contains noise, called noise suppression, noise removal, speech enhancement, etc., and many methods have been proposed. There is. Recently,
As a noise suppression method based on a new concept, a method called the Spectral Mapping method that removes noise from the standpoint of signal detection based on the known mapping relationship between the noise-superimposed signal space generated by using vector quantization and the noise-free signal space. However, the paper “Bi
ing−Hwang Juang, LRRabiner, Signal Restoration by
Spectral Mapping ", 1987 IEEE INTERNATIONAL CONFERE
NCE ON ACOUSTICS, SPEECH, & SIGNAL PROCESSING Volum
e 4 PP.6.6.1-6.6.4 April 1987, Dallas ”(hereinafter, this paper is referred to as reference [1]) and is effective as a noise suppression method for voice transmission.

FIG. 2 can be considered as an example of the configuration of a voice recognition device incorporating a noise removal circuit by the Spectral Mapping method.
Although there are various recognition methods in the recognition device, a discrete word recognition device having a template for each word and using DP (Dynamic Programming) matching will be described as an example.

In FIG. 2, 1 is an input end of an audio signal, 2 is an input audio signal, 3 is an analysis circuit for acoustically analyzing the input audio signal 2, 4
Is a feature vector time series, 5 is a vector quantizer that vector-quantizes the feature vector time series 4 output from the analysis circuit with the noise-added codebook 6, and 7 is a label candidate time series composed of labels representing codewords. , 8 is an inverse vector quantizer for inverse vector quantizing the label candidate time series 7 which is the output of the vector quantizer 5 with the non-soundless codebook 9, 10 is a noiseless feature vector time series, and 11 is the above It is a noise elimination circuit composed of a vector quantizer 5, a noise-added codebook 6, an inverse vector quantizer 8 and a noiseless codebook 9 shown in the document [1]. 12 is a noise-free feature vector time series 10
And a pattern matching circuit that outputs a matching distortion 16 by performing DP matching with a reference pattern 15 represented by a noise-free feature vector time series output from the template memory 14, and 13 is a reference pattern designation and recognition in pattern matching. A recognition control circuit that outputs results, 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, 19 is a pattern matching circuit 12, template memory 14, and recognition control circuit A recognition processing circuit composed of 13 and 13.

Next, the operation will be described.

In the noise elimination circuit 11, the noise-free codebook 9 is configured by using a feature vector of speech without noise superimposed as a codeword, and the noise-added codebook 6 inputs a noise-superimposed input for each codeword of the noise-free codebook 9. A process that is similar to the noise aspect of a voice signal, for example, noise is added in the time waveform domain so that the same signal-to-noise ratio as the noise-superimposed input voice signal 2 is obtained, and this is reanalyzed to add noise It is composed of a code word generated by converting it into a feature vector. The input voice signal 2 input to the input terminal 1 and on which noise is superimposed is acoustically analyzed by the analysis circuit 3, and is a feature vector time series 4 {X (n) | n =
It is output as 1,2, ..., N}. Here, N represents the number of feature vectors. In the noise elimination circuit 11, the vector quantizer 5 receives the feature vector time series 4 as an input, and the noise addition codebook 6 for X (n) corresponding to an arbitrary vector number n.
, And the labels of the code words from the largest likelihood to the L-th rank are label candidates {m _i (n) | i = 1,
2, ..., L}, and for n = 1,2, ..., N, the label candidate time series 7 is {M (n) | n = 1,2, ..., N} (where M (n ) = {M _i (n) | 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 (n) | n =
1,2, ..., N} (where M (n) = {m _i (n) | i = 1,2,
, L}) for any n of the label candidates M (n) =
{M _i (n) | i = 1,2, ..., L} is inverse vector quantized by the noiseless codebook 9 and the noiseless feature vector candidate {Z _i (n)
| i = 1,2, ..., L} is calculated, and Y (n) is calculated as an average vector of the L noise-free feature vector candidates {Z _i (n) | i = 1,2, ..., L}. .

That is, Here, y _t (n) and z _{t, i} (n) are respectively Y
(N) and Z _i (n) are the t-th dimension components.

In the recognition processing circuit 19, the template memory 14 is a reference pattern 15 designated by the address data 17 output from the recognition control circuit 13 {T (k) | k = 1,2, ..., K} (K is a feature vector Number) of the pattern is sent to the pattern matching circuit 12.
The pattern matching circuit 12 is the above reference pattern 15 {T
(K) | k = 1,2, ..., K} and the noise-free feature vector time series 10 which is the output of the noise elimination circuit 11 {Y (n)
DP matching with | n = 1,2, ..., N} is performed. The recurrence formula of DP matching is, for example, as follows.

Here, D (k, n) is the distortion between the feature vector T (k) and the feature vector Y (n). (Equation 2) is an example of the case of DP matching without inclination limitation. This g (K, N) obtained in this recurrence formula (Formula 2) is normalized by dividing it by the urbanization distance (K + N), and the matching distortion 16
Output as. The recognition control circuit 13 uses the address data 17
The reference pattern designated by is changed sequentially, and the label of the reference pattern that minimizes the matching distortion is output as the recognition result 18 from the matching distortion 16 output by the pattern matching circuit 12 for each reference pattern.

[Problems to be Solved by the Invention]

The mapping relationship between the noise-superimposed signal space and the noise-free signal space, which forms the basis of noise removal by the Spectral Mapping method, is obtained by adding the noise-free codebook 9 with a process equivalent to noise superposition to create the noise-added codebook 6. It is formed. However, since the actual noise has a statistical variance, an error occurs in the mapping relationship due to the difference between the noise equivalently added in the generation of the mapping relationship and the noise superimposed on the input voice signal 2. Further, as the level of noise superimposed on the input voice signal 2 rises, the spectrum envelope of the input voice signal 2 is smoothed. Therefore, the phonological feature included in the feature vector time series 4 output from the analysis circuit 3 disappears, and the above-mentioned mapping error significantly increases. For example, when the number L of label candidates in the noise removing circuit 11 is set to 1, when a mapping error due to the variance of noise occurs, the distortion due to the mapping error is directly reflected in the matching distortion 16 in the pattern matching.
The recognition performance drops sharply as the level of superposed noise increases. Moreover, if L is set to a number of 2 or more to increase the number of candidates,
The probability that a candidate with a correct mapping is included in a plurality of candidates increases, but it is not known which candidate among the plurality of candidates is a candidate with a correct mapping. Therefore, the above-mentioned document [1] uses an improvement measure that outputs the average of a plurality of candidate feature vectors. However, this averaging process causes distortion because the component of the erroneous candidate feature vector is mixed into the originally selected candidate feature vector, and the recognition performance is degraded. Therefore, the method of the above-mentioned document [1] has an effect of improving the auditory signal-to-noise ratio in voice transmission, but has no effect when applied to voice recognition. As described above, the conventional voice recognition device having a noise removal function has a problem that the noise removal process causes distortion in the feature vector of the voice, and the recognition rate of the voice decreases under high noise.

The present invention has been made to solve the above problems, and obtains a voice recognition device that does not distort a voice feature vector by noise removal processing and that does not reduce the voice recognition rate even under high noise. The purpose is to

[Means for Solving the Problems]

A speech recognition apparatus according to the present invention includes a noiseless codebook 9 in which a feature vector of speech without noise is used as a codeword.
A noise-added codebook 6 generated by subjecting each codeword of the noise-free codebook 9 to a process equivalent to the noise weight, and a feature vector time series of an input voice signal on which noise is superimposed, Vector quantizing means (ambiguous vector quantizer 20) for outputting a plurality of label candidate time series composed of labels representing codewords according to the above, and this vector quantizing means (ambiguity vector quantizer 20) An inverse vector quantization means (inverse fuzzy vector quantizer 22) for inverse vector quantizing a plurality of label candidate time series which are output signals in accordance with the noiseless codebook 9 and outputting a plurality of noiseless feature vector candidate time series; , All the noise-free feature vectors by inputting a plurality of noise-free feature vector candidate time series which are output signals of the inverse vector quantizer (inverse ambiguous vector quantizer 22) A multi-vector recognition processing means (multi-vector recognition processing circuit 26) for selecting the optimum noise-free feature vector time series based on the reference pattern giving the maximum likelihood for the complementary time series and performing the voice recognition processing. It is characterized by.

[Action]

The vector quantizer (ambiguous vector quantizer 20) vector-quantizes the feature vector time series of the input speech signal on which noise is superimposed based on the noise-added codebook 6, and inverse vector quantizes a plurality of label candidate time series. Means (inverse fuzzy vector quantizer 22). The inverse vector quantizer (inverse ambiguous vector quantizer 22) inverse vector quantizes 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. It is given to the vector recognition processing means (multiple vector recognition processing circuit 26). The multi-vector recognition processing means (multi-vector recognition processing circuit 26) is optimal based on the reference pattern that gives the maximum likelihood for all the noise-free feature vector candidate time series from the input plurality of noise-free feature vector candidate time series. A noise-free feature vector time series is selected and speech recognition processing is performed.

Example of Invention

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

In FIG. 1, 20 is a vector quantizing means for vector-quantizing the feature vector time series 4 which is the output signal of the analyzing circuit 3 using the noise addition codebook 6 and outputting a plurality of label candidate time series 21. Fuzzy vector quantizer,
Reference numeral 22 denotes an inverse fuzzy vector quantizer as an inverse vector quantizer that inverse vector quantizes a plurality of label candidate time series 21 with the noiseless codebook 9 and outputs a plurality of noiseless feature vector candidate time series 23. A noise-free multiple vector generation circuit composed of a noise-added codebook 6, a noiseless codebook 9, an ambiguous vector quantizer 20 and an inverse ambiguous vector quantizer 22;
25 is a multiple vector pattern matching circuit that inputs a plurality of noise-free feature vector candidate time series 23 and performs DP matching with reference pattern 15 and outputs matching distortion 16.
Is a multiplex vector recognition processing circuit as multiplex vector recognition processing means composed of multiplex vector pattern matching circuit 25, template memory 14 and recognition control circuit 13.

This speech recognizer uses fuzzy vector quantization (Fuzzy Vect
or Quantization), a new method that simultaneously performs noise removal and pattern recognition of the input signal is adopted, and the feature vector of the input speech signal on which noise is superimposed is input to generate multiple noise-free feature vector candidate time series 23. A noiseless multiplex vector generation circuit 24 for outputting and a multiplex vector recognition processing circuit 26 for performing recognition with a plurality of noiseless feature vector candidate time series 23, which are output signals of the noiseless multiplex vector generation circuit 24, as inputs. As described above, the noiseless multiplex vector generation circuit 24 performs noiseless codebook 9 in which the feature vector of speech without noise is used as a codeword, and noise superimposition in each codeword of the noiseless codebook 9. The noise-added codebook 6 generated by equivalent processing and the feature vector time series 4 of the input speech signal 2 on which noise is superimposed are vector-quantized in accordance with the noise-added codebook 6 to obtain a plurality of label candidate time series 21.
Ambiguous vector quantizer 20 and a plurality of label candidate time series 21, which are output signals of the ambiguous vector quantizer 20, are inverse vector quantized according to the noiseless codebook 9 to obtain a plurality of noiseless feature vector candidates. It is composed of an inverse ambiguity vector quantizer 22 that outputs a time series 23.

The multiple vector recognition processing circuit 26 is {Z (n) | n = 1,2, ..., N} (however, a plurality of noise-free feature vector candidate time series 23 which are output signals of the deambiguity vector quantizer 22. Z
(N) = {Z _i (n) | i = 1,2, ..., L}), and the noise-free feature vector candidate Z (n) = {Z
Among L candidates of _i (n) | i = 1,2, ..., L}, a candidate that maximizes the likelihood of the reference pattern is selected as an optimal noiseless feature vector for each reference pattern. As a result, the reference pattern that gives the maximum likelihood for all time series up to n = N is finally determined as the recognition category, and at the same time, the processing for determining the optimum noise-free feature vector time series is performed.

The operation of this embodiment will be described.

The ambiguity vector quantizer 20 receives {X (n) | n = 1,2, ..., N}, which is the feature vector time series 4 of the input speech, as an input and outputs the feature vector X (n) for any n. Likelihoods with all codewords in the noise-added codebook 6 are calculated, and the L-th is calculated from the largest likelihood.
Labels of code words up to the rank are label candidates {m _i (n) | i =
1,2, ..., L}, this is obtained for n = 1,2, ..., N and a plurality of label candidate time series 21 {M (n) | n = 1,2, ..., N} (where M ( n) = {m _i (n) | i = 1,2, ..., L}). The inverse ambiguity vector quantizer 22 includes {M (n) | n = 1,2, ..., N} (where M is a plurality of label candidate time series 21).
Label candidates M (n) = {m _i (n) | i = 1,2, ... Corresponding to any n in (n) = {m _i (n) | i = 1,2, ..., L}) ,
L} is inverse-vector quantized by the noiseless codebook 9, and a plurality of noiseless feature vector candidates {Z _i (n) | i = 1,2, ..., L}
, N is obtained for n = 1, 2, ..., N, and is a plurality of noise-free feature vector candidate time series 23 {Z (n) | n = 1, 1.
2, ..., N} (where Z (n) = {Z _i (n) | i = 1,2, ...,
Output as L}).

The multi-vector pattern matching circuit 25 is a plurality of noise-free feature vector candidate time series 23 which are output signals of the deambiguity vector quantizer 22 {Z (n) | n = 1,2, ...,
N} (however, Z (n) = {Z _i (n) | i = 1,2, ..., L})
, And input the reference pattern 15 {T (k) | k = 1,2, ..., K} (K
Performs DP matching (equation 2) with the number of feature vectors). However, 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 by the following equation.

D (k, n) = min (d (T (k), Z _i (n)) 1 ≦ i ≦ L (Equation 3) where d (*, *) represents the distortion between the feature vectors. , The optimum feature vector for the reference pattern 15 is selected from a plurality of candidates.As in the conventional example, g (K, N) obtained in the recurrence formula (Formula 2) is converted into the urbanization distance ( It is normalized by dividing by K + N) and is output as matching distortion 16. By the principle of DP matching, if the partial distortion of (Equation 3) is minimized, the matching distortion of the entire pattern is minimized. Sequentially changes the reference pattern in the template memory 14 specified by the address data 17,
The matching distortion 16 output by the pattern 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. For a reference pattern that minimizes the matching distortion 16, a vector time series that is optimally selected from among a plurality of noise-free feature vector candidate time series 23 is a speech signal in which noise is correctly removed from noise-superimposed input speech. It is approximated to 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. The DP matching according to (Equation 2) is given as an example, and does not limit the scope of application of the present invention.

As described above, in the noiseless multiple vector generation circuit 24, the feature vector time series 4 of the input voice on which noise is superimposed is ambiguous by providing redundancy to the plurality of noiseless feature vector candidate time series 23 including the correct noise removal signal. In the multi-vector recognition processing circuit 26, the recognition processing is performed by selecting the noise-free feature vector time series that automatically minimizes the mapping distortion by the processing that maximizes the likelihood by using the noise-free reference pattern. Done.

Therefore, in this embodiment, the problem of the above-mentioned conventional device, the distortion due to the mapping error and the distortion caused by averaging the feature vectors of a plurality of candidates are eliminated, and the reduction of the recognition rate due to them does not occur. .

As described above, the present invention has been described by taking the discrete word voice recognition device that registers the template of the word unit as an example, but the recognition unit is not limited to the word, and CV (consonant-vowel),
VCV (vowel-consonant-vowel), CVC (consonant-vowel-consonant),
It may be an arbitrary unit such as a monosyllabic or a morpheme, and as long as template registration is performed in time series of feature vectors,
The same effect as that of the above embodiment is obtained. HMM (Hidden Markov Mod) other than the recognition method by template matching
Even in statistical recognition methods such as el), a transition probability model is used instead of a reference pattern, and a noise-free feature vector that maximizes the generation probability is selected from each noise-free feature vector candidate for each transition probability model. The same effect can be obtained with. Furthermore, even for continuously uttered voices other than discrete word voices, a segmentation circuit that cuts out the continuous utterances into registered voice units is installed in the preceding stage of this device, and the cutout voices are input to this device. Applicable. Needless to say, the implementation means of the present invention is not limited to dedicated hardware, and can be implemented by software processing in a general-purpose computer or signal processor. Furthermore, the present invention can be extended and applied to a recognition device for acoustic signals other than voice, image signals, and character / graphic signals.

〔The invention's effect〕

As described above, according to the present invention, a vector quantization means for vector-quantizing a feature vector time series of an input speech signal on which noise is superimposed according to a noise-added codebook, and outputting a plurality of label candidate time series, and this vector quantization Inverse vector quantization means for inverse vector quantizing a plurality of label candidate time series as output signals of the means in accordance with the noiseless codebook and outputting a plurality of noiseless feature vector candidate time series, and output of the inverse vector quantization means Input a plurality of noise-free feature vector candidate time series, which are signals, and give the maximum likelihood for all noise-free feature vector candidate time series. Since it is configured with multiple vector recognition processing means for performing recognition processing, distortion with reference patterns is minimized during pattern matching. It is possible to select the optimum noise-free feature vector time series from among a plurality of noise-free feature vector candidates on the basis of the fact that the distortion of the voice feature vector due to the denoising process does not occur. The effect is that the recognition rate is extremely high even under high noise.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a voice recognition device according to an embodiment of the present invention, and FIG. 2 is a block diagram showing the configuration of a conventional voice recognition device. 6 ... Noise-added codebook, 9 ... No-noise codebook, 20 ...
Ambiguous vector quantizer (vector quantizer), 22 ...
Inverse fuzzy vector quantizer (inverse vector quantizer), 26
... Multi-vector recognition processing circuit (multi-vector recognition processing means).

Claims (1)

[Claims] 1. A speech recognition apparatus for recognizing an input speech having a noise removal function for suppressing noise superimposed on an input speech signal, wherein a noiseless code having a feature vector of speech on which no noise is superimposed as a code word. Book, a noise-added codebook generated by performing a process equivalent to noise superposition on each codeword of this noise-free codebook, and a feature vector time series of an input voice signal on which noise is superposed, according to the above noise-added codebook. A vector quantizing means for outputting a plurality of label candidate time series composed of labels representing quantized code words, and a plurality of label candidate time series as an output signal of the vector quantizing means are inversed according to the noiseless codebook. Inverse vector quantization means for vector-quantizing and outputting a plurality of noise-free feature vector candidate time series, and a plurality of noise-free output signals of the inverse vector quantization means A multi-vector for inputting the feature vector candidate time series and giving the maximum likelihood for all noise-free feature vector candidate time series, selecting the optimum noise-free feature vector time series based on the reference pattern, and performing speech recognition processing. A speech recognition apparatus comprising: a recognition processing unit.