JP2961797B2 - Voice recognition device - Google Patents

Abstract

PURPOSE:To improve the accuracy of recognition by setting the degree of contribution at every identification symbol in the case of generating a symbol model and calculating likelihood based on a contribution degree information dictionary at every identification symbol of an acoustic feature parameter. CONSTITUTION:An input sound 1 is converted to a sound feature sequence 3 in a sound analysis part 2. In the case of generating the symbol model, the input sound is inputted to a symbol model generation part 6 as a feature sequence 4 for learning and the symbol model 7 is generated. At this time, a contribution degree setting part 16 by symbols sets a feature contribution degree 17 for symbol model generation according to the contribution degree information dictionary 15. Next, a recognition object model generation part 8 generates a recognition object model 10 by linking the symbol models while referring to a recognition object dictionary 9. In a recognition processing, the sequence 3 is inputted to a likelihood calculation part 11 and the likelihood 12 is calculated based on a feature contribution degree 18 for the likelihood calculation from the setting part 16. A sellective part 13 selects and outputs the recognition object model having the highest likelihood. Thus, the accuracy of recognition is improved while considering the difference of an effect to recognition performance between the feature parameters concerning the respective identification symbols.

Description

Description: TECHNICAL FIELD The present invention relates to a speech recognition device for recognizing speech.

[Prior Art] Conventionally, as an apparatus of this type, for example, there is one as shown in FIG. This figure shows Kiki Maruyama, Toshiyuki Hanazawa, Go Kawabata,
Illustration of the contents described in Kiyohiro Kano, "Recognition of English Word Speech Using HMM Phonetic Connection Learning" (Proceedings of the Acoustical Society of Japan, published March 14, 1989, 1-6-22) It is.

In the figure, (1) is a speech signal, (2) is a speech analysis unit, (3) is a speech feature sequence, (4) is a learning feature sequence,
(5) is an input feature sequence, (6) is a symbol model generation unit,
(7) is a symbol model, (8) is a recognition target model generation unit,
(9) is the recognition target dictionary, (10) is the recognition target model, (1)
1) is the likelihood calculator, (12) is the likelihood, (13) is the selector, (1)
4) is the recognition result.

The input speech (1) is converted into a speech feature sequence (3) by a plurality of acoustic feature parameters in a speech analysis unit (2). When generating a symbol model, the speech feature sequence is input to the symbol model generation unit (6) as a learning feature sequence (4). The symbol model generation unit generates a symbol model (7) representing a category corresponding to an identification symbol given to a phoneme unit / phoneme / phoneme sequence or a syllable speech description unit from the learning feature sequence. The recognition target model generation unit (8) generates a recognition target model (10) by connecting the symbol models according to the identification symbol string representing the recognition target word described in the recognition target dictionary (9). When performing the recognition process, the speech feature sequence is input to the likelihood calculation unit (11) as an input feature sequence (5). The likelihood calculation unit calculates the likelihood (12) for the input feature sequence for each recognition target model generated as described above. The selecting unit (13) selects a recognition target model having the highest likelihood among them and outputs this as a recognition result (14).

Next, details of the above-described conventional example will be described in accordance with the above configuration. Here, it is assumed that the acoustic feature parameters are LPC cepstrum, three types of feature vectors of cepstrum difference and power, the identification symbol is a phoneme, and English word speech is recognized. In addition, vector quantization (Vector Quantizat
HMM (Hidden Markov M) based on ion, VQ
odel). This is very simply as follows.

Vector quantization (VQ) selects a nearest neighbor vector from a set of vectors prepared in advance for an input vector (called a VQ codebook), and determines that the nearest neighbor vector represents the input vector well. This is done by outputting a number (called a VQ label). Discrete distribution HMM
Is a probabilistic model that is composed of several states and transitions between them, and expresses a label sequence, using the probabilities of transitions between states and the output probability distribution of labels output at the transition as parameters. In speech recognition using this, the occurrence probability of a VQ label sequence obtained by VQ for the acoustic feature parameter vector of each input speech frame is compared among a plurality of HMMs,
This is done by selecting the HMM that has the highest probability.

In a discrete distribution HMM, the local likelihood of a model with respect to an input frame is represented by an output probability of an input VQ label.
At this time, in the conventional example, the output probability for each feature of the HMM representing the state transition i for each of the VQ label sets (L1, L2, L3) based on the three types of acoustic feature parameters of the first frame is represented by b ^L1 _il. , b ^L2 _il and b ^L3 _il , the overall output probability B _il is given by equation (1).

B _il = b ^L1 _il × b ^L2 _il × b ^L3 _il (1) The likelihood of the input speech feature sequence for the recognition target HMM is
When the above-mentioned sequence makes an optimal state transition with respect to this HMM, it is substantially equal to the sum of output probabilities obtained from each input frame and the corresponding state transition.

Hereinafter, detailed operations of the speech analysis unit, the symbol model generation unit, and the likelihood calculation unit will be described.

[A] Voice analysis unit FIG. 3 shows details of the voice analysis unit. In the figure,
(1) is an audio signal, (2) is an audio analysis unit, (3) is an audio feature sequence (VQ label sequence), (A1) is an audio analysis unit, and (A2)
Is a feature vector sequence, (A3, A4, A5) is a VQ codebook (three types), and (A6) is a vector quantization unit. The input audio signal (1) is converted into a feature vector sequence (A2) composed of three types of acoustic feature vectors by an acoustic analysis unit (A1). The vector quantization unit (A6) divides the feature vector sequence into three using three VQ codebooks (A3), (A4), and (A5).
Into a speech feature sequence (3) consisting of a set of two VQ labels.

[B] Symbol model generator FIG. 4 shows details of the symbol model generator. In the figure, (4) is a learning feature sequence (word VQ label sequence),
(6) is a symbol model (phoneme HMM) generation unit, (7) is a symbol model (phoneme HMM), (B1) is an initial phoneme HMM, (B2) is a pre-learning phoneme NMM, (B3) is a phoneme HMM connection unit, (B4) is a word dictionary for learning, (B5) is a word HMM before learning, (B6) is a word HMM learning unit, (B7) is a word HMM after learning, (B8) is a word HMM decomposition unit, and (B9) is learning. Post-phonological HMM.

 The processing procedure is as follows.

[1] Let the initial phoneme HMM (B1) be the pre-learning phoneme HMM (B2).

[2] Based on learning word dictionary (B4), pre-learning phoneme HM
M is connected by the phoneme HMM connection part (B3), and the pre-learning word HMM (B5)
Generate In the HMM, models need only be connected. Also, if the phonemes required for recognition are included,
The learning word dictionary may be different from the dictionary used for recognition.

[3] The word HMM learning unit (B6) presents a learning feature sequence (word VQ label sequence) (4) to the pre-learning word HMM, and performs model learning. The output probability of the HMM used at this time is obtained by equation (1).

[4] The post-learning word HMM (B7) generated in [3] is decomposed in the word HMM decomposing unit (B8), and the post-learning phoneme HMM (B
9) Generate

[5] The phoneme HMM after learning is used as a pre-learning phoneme, and after [2] is repeated a required number of times, the obtained phoneme HMM is used as a symbol model (7) used for recognition.

[C] Likelihood calculation unit The likelihood (12) of the input feature sequence (5) with respect to the generated recognition target (word) HMM (10) is obtained. At this time, H
The output probability of the MM is obtained by the aforementioned equation (1).

[Problems to be Solved by the Invention] The effect of the acoustic feature parameter on the recognition performance differs depending on the type and also on the category represented by the identification symbol. Therefore, it is possible to improve the recognition performance by considering this point. it can. However, the conventional device treats all acoustic feature parameters equally as can be seen from equation (1), and has a problem that the difference in the effect on recognition performance between them is not reflected in the device.

The present invention has been made to solve such a problem, and an object of the present invention is to provide a speech recognition device having high recognition accuracy by considering a difference in effect on recognition performance between acoustic feature parameters for each identification symbol. .

[Means for Solving the Problems] A speech recognition apparatus according to the present invention provides a symbol model that represents a category corresponding to an identification symbol given to a speech description unit by a plurality of types of acoustic feature parameters, by using a plurality of recognition target objects. Speech recognition for generating a recognition target model by concatenating voices according to a recognition target dictionary represented by the series of identification symbols, and outputting, as a recognition result, one of the recognition target models having the highest likelihood with respect to the input voice as a recognition result The apparatus is configured to determine the contribution of each of the plurality of types of acoustic feature parameters to the likelihood obtained by the symbol model when the symbol model is generated by the symbol model generation means and when the likelihood calculation means calculates the likelihood. A symbol-by-symbol contribution setting means for individually setting each symbol, and the symbol-by-symbol contribution setting means performs the contribution setting. And a contribution information dictionary in which information regarding the contribution of the plurality of types of acoustic feature parameters is described for each of the identification symbols.

[Operation] In the present invention, in the contribution information dictionary, information on the contribution setting of each of the plurality of types of acoustic feature parameters with respect to the likelihood obtained by the symbol model is described for each of the identification symbols. Contribution setting means, based on the contribution information dictionary, individually calculates the contribution at the time of symbol model generation by the symbol model generation means and at the time of likelihood calculation by the likelihood calculation means for each of the identification symbols. Set.

[Example] Here, the acoustic feature parameter is defined as a static feature (LPC cepstrum) and a dynamic feature (cepstrum regression coefficient).
A case is described in which the type feature vector and the identification symbol are phoneme segments, and the word speech is recognized by the discrete distribution HMM. When the conventional method is used under these conditions, the speech analysis unit uses two types of acoustic feature vectors (static features and
A feature vector sequence composed of dynamic features) is converted into a VQ label sequence in accordance with two VQ codebooks prepared in advance, and HMM learning and input speech recognition are performed using the sequence as a speech feature sequence.

When a phoneme segment is used as an identification symbol representing a recognition target, a stationary part and a transient part of speech can be treated independently.
When static and dynamic features are used as acoustic features, static features are recognized in stationary parts with small temporal feature fluctuations, and dynamic features are recognized in transient parts with large temporal feature fluctuations. It is thought that they carry useful information. However, when the output probability is calculated based on the equation (1) by the conventional device, the contributions of these two types of features to the likelihood are equal, and the difference in the effect on the recognition performance between the feature parameters is not reflected on the device. . Therefore,
The recognition performance is improved by increasing the contribution of the static feature in the stationary part speech unit model and the dynamic feature in the transient part speech unit model.

 FIG. 1 is a block diagram of one embodiment of the present invention.

In the figure, (1) is a speech signal, (2) is a speech analysis unit, (3) is a speech feature sequence, (4) is a learning feature sequence,
(5) is an input feature sequence, (6) is a symbol model generation unit,
(7) is a symbol model, (8) is a recognition target model generation unit,
(9) is the recognition target dictionary, (10) is the recognition target model, (1)
1) is the likelihood settlement section, (12) is the likelihood, (13) is the selection section, (1)
4) is the recognition result, (15) is the contribution information dictionary, (16) is the contribution setting unit for each symbol, (17) is the acoustic feature contribution for symbol model generation, and (18) is the acoustic for likelihood calculation. The characteristic contribution.

Hereinafter, the operation of this embodiment will be described. The input speech (1) is converted into a speech feature sequence (3) by a plurality of acoustic feature parameters in a speech analysis unit (2). When generating a symbol model, the speech feature sequence is input to the symbol model generation unit (6) as a learning feature sequence (4). The symbol model generation unit generates, from the learning feature sequence, a symbol model (7) representing a category corresponding to an identification symbol assigned to a speech description unit such as a phoneme segment or a phoneme. At this time, the symbol-specific contribution setting unit (16) sets the characteristic contribution for symbol model generation (17) for each identification symbol in accordance with the contribution information dictionary (15), and the symbol model is generated according to the contribution. The recognition target model generation unit (8) is a recognition target dictionary (9)
The recognition target model (10) is generated by connecting the symbol models according to the identification symbol string representing the recognition target word described in (1). When performing the recognition process, the speech feature sequence is input to the likelihood calculation unit (11) as an input feature sequence (5). The likelihood calculation unit calculates the likelihood (12) for the input feature sequence for each recognition target model generated as described above.
Is calculated. At this time, the symbol-by-symbol contribution setting unit sets the likelihood calculation characteristic contribution (18) for each identification symbol according to the contribution information dictionary. The selection unit (13) selects the recognition target model that has obtained the highest likelihood, and outputs this as the recognition result (14).

Hereinafter, processing in the symbol-specific contribution degree setting unit in this embodiment will be described. As described above, in the conventional apparatus, the output probabilities obtained by the HMM for a plurality of types of acoustic feature parameters can be obtained by the equation (1). In this embodiment, the output probabilities are obtained by the following equation (2). 2 of the first frame
The output probabilities for each feature of the HMM representing the state transition i for each pair of VQ labels s (static feature) and d (dynamic feature) are b ^s _il and b ^d _il , respectively, and the overall output probability B _il the _{^{B il = (b s il)}} X × (b d il) Y ...... (2) X = W is, Y = W id (W is, W id is weight) and. At this time, the weight W for the output probabilities for the two features
_{It IS,} W _id makes it possible to set the contribution for each phoneme to the likelihood of the two characteristic parameters. Thus, the symbol-specific contribution setting unit sets these weights for each phoneme segment, and uses these weights when generating the symbol model and when calculating the likelihood at the time of recognition.
Here, two sets of weights W _cs , W _cd and
Set W _ts and W _td .

The setting of the contribution degree of the static feature and the dynamic feature in the stationary part speech segment and the transient part speech segment in the form described above is realized as follows. First, the contribution information dictionary describes, for each phoneme segment used for recognition, whether it is a stationary part or a transient part. The symbol-by-symbol contribution degree setting unit sets the set of weights (W _is , W _id ) to be used when obtaining the output probability according to equation (2).
At this time, from the contribution information dictionary, the set of weights is determined as W _cs , W _cd (stationary part) and W _ts , W _td depending on whether the phoneme segment is a stationary part or a transient part.
(Transient section). Here, the weight is W _cs ≧
It is assumed that the condition of W _cd , W _ts ≦ W _td is satisfied. When the output probability is calculated by the equation (2) at the time of model generation by the symbol model generation unit and at the time of likelihood calculation by the likelihood calculation unit, the stationary unit speech unit HMM uses the above-mentioned weight setting for static features. Emphasizing the output probability (W _cs ≧ W _cd )
M emphasizes the output probability for dynamic features (W _ts ≦
_Wtd ) Can.

Here, for the order of B _il in common regardless of speech segment, there needs to be W is ₊ W _id = C _{(constant) (W is, W id ≧} 0). Note that the weights may not be common between the generation of the symbol model and the calculation of the likelihood.

In this embodiment, a case has been described in which two types of feature vectors, a static feature and a dynamic feature, are used as acoustic feature parameters, and a discrete HMM is learned from word speech using phonemic segments as identification symbols to recognize words. However, the acoustic feature parameter may be power or the number of zero crossings, the identification symbol may be a phoneme or a syllable, and the speech used for recognition may be sentence speech. Further, the recognition method may be a DTW (Dynamic Time Warping) method using a template in units of categories such as phonemes.
That is, these conditions in this embodiment do not limit the present invention.

[Effects of the Invention] As described above, according to the present invention, a symbol model expressing a category corresponding to an identification symbol given to a description unit of a speech by a plurality of types of acoustic feature parameters is obtained by converting a plurality of recognition target speeches into A speech recognition device that generates a recognition target model by linking according to a recognition target dictionary represented by a sequence of identification symbols, and outputs a recognition target model that obtains a value with the highest likelihood for input speech as a recognition result, At the time of symbol model generation by the symbol model generation means and at the time of likelihood calculation by the likelihood calculation means, the contribution of each of the plurality of types of acoustic feature parameters to the likelihood obtained by the symbol model is determined for each of the identification symbols. Symbol-specific contribution setting means to be individually set, and the symbol-specific contribution setting means refers to when performing the contribution setting. A contribution information dictionary in which information on the contributions of a plurality of types of acoustic feature parameters is described for each of the identification symbols is provided, so that the acoustic feature parameters of the categories represented by the identification symbols are considered. A highly accurate speech recognition device can be provided by performing emphasis for each identification symbol.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a speech recognition apparatus according to an embodiment of the present invention, FIG. 2 is a block diagram showing a speech recognition apparatus according to a conventional example, and FIG. 3 is a speech analysis unit (2) shown in FIG. FIG. 4 is a block diagram illustrating the symbol model generator (6) of FIG. In the figure, (1) is an audio signal, (2) is an audio analysis unit,
(3) is a speech feature sequence, (4) is a learning feature sequence,
(5) is an input feature sequence, (6) is a symbol model generation unit,
(7) is a symbol model, (8) is a recognition target model generation unit,
(9) is the recognition target dictionary, (10) is the recognition target model, (1)
1) is the likelihood calculator, (12) is the likelihood, (13) is the selector, (1)
4) is the recognition result, (15) is the contribution information dictionary, (16) is the contribution setting unit for each symbol, (17) is the acoustic feature contribution for symbol model generation, and (18) is the acoustic for likelihood calculation. The characteristic contribution. (A1) is an acoustic analysis unit, (A2) is a feature vector sequence, (A3), (A4), and (A5) are VQ codebooks, (A
6) is the vector quantizer, (B1) is the initial phoneme HMM, (B2)
Is the pre-learning phoneme HMM, (B3) is the phoneme HMM connection part, (B4) is the learning word dictionary, (B5) is the pre-learning word HMM, and (B6) is the word H
The MM learning unit, (B7) is a word HMM after learning, (B8) is a word HMM decomposition unit, and (B9) is a phoneme HMM after learning. In the drawings, the same reference numerals indicate the same or corresponding parts.

Continuation of the front page (56) References JP-A-63-158596 (JP, A) JP-A-61-177493 (JP, A) JP-A-2-29799 (JP, A) JP-A-63-15299 (JP) JP-A-60-182496 (JP, A) JP-A-58-145999 (JP, A) Proceedings of the Acoustical Society of Japan October 1988 2-P-21 p. 243-244 Proceedings of the Acoustical Society of Japan March 1989 1-6-21 p. 41-42 Proceedings of the 1990 IEICE Spring Conference Annual Meeting A-230 p. 1-230 Proceedings International Conference on Acoustics, Speach and Signal Processing, 1987, Vol. 2 p. 1163-1166, Proceedings International Conference on Acoustics, Speeche and Signal Processing, 1989, Vol. 1 p. 298-301 Electronic Letters 6th July 1989, Vol. 14, p. 918-919 IEICE Technical Report Vol. 88, No. 329, SP88-106, pp. 1-8 (1988/12/16) IEICE Technical Report Vol. 88, No. 329, SP88-107, pp. 9-15 (1988/12/16) IEICE Technical Report Vol. 87, No. 299, SP87-97, pp. 1-6 (1987/12/18) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00 535 G10L 5/06 G10L 7/08 JICST Science and Technology Reference file

Claims (1)

(57) [Claims] 1. A symbol model generating means for model-representing a category corresponding to an identification symbol given to a description unit of a speech by a plurality of types of acoustic feature parameters, and a recognition unit for representing a plurality of recognition target speeches by a series of the identification symbols. Target dictionary,
A recognition target model generating means for generating a recognition target model expressing the recognition target voice by concatenating the symbol models according to the recognition target dictionary, and a likelihood of the input voice for the recognition target model generated by the recognition target model generation means. Speech recognition apparatus comprising: a likelihood calculating means for obtaining a degree, and a selecting means for selecting a recognition target model obtaining the highest value among the likelihoods obtained by the likelihood calculating means, and outputting the selected model as a recognition result. In the above, when the symbol model is generated by the symbol model generating means and the likelihood is calculated by the likelihood calculating means, the contribution of each of the plurality of types of acoustic feature parameters to the likelihood obtained by the symbol model is determined by the identification symbol. A symbol-specific contribution degree setting unit that is individually set for each, and this symbol-specific contribution degree setting unit refers to when performing the contribution degree setting, Serial plurality of types of speech recognition device, characterized in that it comprises the contribution information dictionary with information about the contribution of the acoustic feature parameters for each of the identification mark.