JP2000099071A - Speech recognition device and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the recognition performance by considering the relationship of acoustic features of a segment and the ones immediately before and after. SOLUTION: Speech to be learned is transformed into acoustic feature parameter. Then, the loci of acoustic parameters are obtained for any one of a section 1 which includes a recognition object phoneme Wi and the tail of a phoneme Wi-1 that is immediately before the phoneme Wi, a section 2 which includes the phoneme Wi and the leading top of a phoneme Wi+1 that is immediately after the phoneme Wi and a section 3 which includes the phoneme Wi, the tail of the phoneme Wi-1 and the phoneme Wi+1 that is immediately after the phoneme Wi. Then, appearance probabilities P(Bi-1, Ai|Wi), P(Ai, Bi+1|Wi) and P(Bi+1, Ai, Bi+1|Wi) are obtained against the phoneme Wi to generate models. Then, inputted voices are transformed into feature parameters during a speech recognition, loci of the section parameters corresponding to the sections of the models for each phoneme are obtained and the likelihood between the loci and each model is obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a speech recognition apparatus and method using a segment model for recognizing speech based on the locus of acoustic feature parameters.

[0002]

2. Description of the Related Art Conventionally, basic units of speech recognition include phoneme units, subword (subword) units, word units, and the like (hereinafter, these are referred to as units). HMM) is widely used as an acoustic model (eg, Seiichi Nakagawa, “Speech Recognition by Stochastic Model”, IEICE, Showa 6)
See issue in July 2013. ). In speech recognition, speech is parameterized at certain fixed time intervals (here, this is called a frame). In the method based on the HMM, assuming that values of parameters between adjacent frames are independent, modeling of speech and calculation of likelihood of a recognition candidate are performed. On the other hand, due to the restriction of the human vocalization mechanism, the feature parameters of the speech are not considered to be independent in adjacent frames. As a model that reinforces this point, a segment model assuming continuity of parameter values in a unit has been proposed (for example, M. Ostendorf et al., “From HMMs to segment models”
: Aubified view of stochastic modeling for speech
recognition ”IEEE Transactions on Speech and Audi
o Processing, SAP-4 (5), pp. 360-378 (1996-9)).

[0003]

In the conventional HMM, the parameter values are assumed to be independent, and the continuity of the trajectory of the parameter cannot be sufficiently handled. In addition, although the segment model up to now has been concerned with the continuity of the parameters inside the unit, it does not deal with the continuity with the parameter values outside the unit (between adjacent units), and the recognition performance is still insufficient. Was not. An object of the present invention is to solve the above problem by considering the continuity of parameter values not only within the segment (unit) but also with an adjacent segment (unit), and efficiently model the problem. An object of the present invention is to provide a speech recognition apparatus and method having a method.

[0004]

According to the present invention, there is provided a speech recognition apparatus using a segment model for analyzing an inputted speech into speech acoustic feature parameters and performing recognition based on information of a locus of the feature parameters. , The section including the end of the segment immediately before the segment to be recognized, the section including the beginning of the immediately following segment, or the end of the immediately preceding segment and the beginning of the immediately following segment The trajectory of the parameter is obtained by combining the feature parameter of the segment, that is, the feature parameter of the segment section including the feature parameter of the segment to be recognized, and the likelihood of the segment based on the information of the trajectory. It is characterized in that speech is recognized using degrees. That is, the appearance probability of the trajectory of the feature parameter including the transition portion with respect to the segment information is obtained in advance as a model, and the likelihood between this model and the trajectory of the feature parameter of the input audio signal is obtained.

According to a second aspect of the present invention, in the first aspect of the invention, the likelihood of the segment is calculated in consideration of the label information of segments before and after the segment in the likelihood calculation of the segment. It is characterized by the following.

[0006]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a range for obtaining a locus of a characteristic parameter which is a main part of the present invention. In FIG. 1, the label of the i-th segment (specifically, phoneme, sub-word (sub-word), word) to be recognized is wi, the label of the preceding segment is wi-1, and the label of the following segment is wi + 1 and each. The trajectories of the characteristic parameters obtained for each frame at the labels wi, wi-1 and wi + 1 of the respective segments are denoted by Ai, Ai-1 and Ai + 1, respectively. In the present invention, if all of the preceding and succeeding segments are used, not only does the processing amount increase, but also the trajectory estimation accuracy decreases, so the transition part of the preceding and following segments, that is, the segment immediately before the segment to be recognized, has a trailing end. Only the leading portion Bi + 1 is considered for the portion Bi-1 and the immediately succeeding segment. Specifically, if the segment is a phoneme, its length is usually 5
It is about 0 to 100 milliseconds, but the transition parts Bi-1 and Bi +
1 is about 10 to 50 milliseconds.

When a parameter trajectory is obtained in a section including the end of the segment immediately before the segment to be recognized, the section 1 in FIG. 1 is obtained. Bi-1,
The probability of occurrence of Ai is represented by P (Bi-1, Ai | wi) or the probability P (Bi-1, Ai | wi) / P (Bi-1 | wi) normalized by the appearance probability of the previous segment. In addition, when the locus of the parameter is obtained in the section including the leading part of the immediately following segment, the section becomes the section 2,
The probability of the locus appearing is P (Ai, Bi + 1 | wi) or the probability P (Ai, Bi + 1 | wi) / P (Bi + 1 | wi) normalized by the appearance probability of the subsequent segment. Represent. When a parameter trajectory is obtained in a section including the end of the immediately preceding segment and the beginning of the immediately following segment, the section 3 is obtained, and the probability of occurrence of the trajectory is represented by P (Bi-1, Ai, Bi). +1 | wi).

On the other hand, with regard to the acoustic segment model depending on the context (for example, phoneme environment), when the trajectory of the parameter is obtained in the section including the end of the segment immediately before the segment to be recognized, the section becomes the section 1. , The probability that the locus appears is P (Bi−1, Ai | wi−1, wi, wi + 1) or the probability P (Bi−1, Ai | wi−1) normalized by the appearance probability of the previous segment. , wi, wi + 1) / P (Bi-1 | w
i-1, wi, wi + 1). In addition, when the locus of the parameter is obtained in the section including the leading part of the immediately following segment, the section becomes the section 2,
The probability of appearance of the trajectory is P (Ai, Bi + 1 | wi-1, Wi, wi + 1) or the probability P (Ai, Bi + 1 | wi-1, normalized by the appearance probability of the subsequent segment. wi, wi + 1) / P (Bi + 1 | w
i-1, wi, wi + 1). When a parameter trajectory is obtained in a section including the last part of the immediately preceding segment and the first part of the immediately following segment, the trajectory becomes Section 3, and the appearance probability of the trajectory is represented by P (Bi-1, Ai, Bi). +1 | wi-1, wi, w
i + 1).

The context-dependent acoustic segment model includes label information of a segment to be recognized,
Only the label information of the segment immediately before or immediately after may be considered. FIG. 2 is a block diagram for creating an acoustic segment model used in this embodiment. The input learning speech data is converted into feature parameters such as cepstrum by the feature extraction unit 12, and the trajectory of each parameter is estimated by the trajectory calculation unit 13 according to the estimated section of the trajectory.
A model creation unit 1 uses a set of these trajectories and label data of input learning speech data (in which utterance contents are described).
In step 4, an acoustic segment model is created and stored in the memory 15.

FIG. 3 is a block diagram of the speech recognition system of this embodiment. The voice input from the input terminal 21 is
The feature extraction unit 22 converts the parameters into feature parameters such as cepstrum and the like.
In step 3, the trajectory of each parameter is estimated. From memory 24,
The likelihood of the recognition candidate generated using the word dictionary 25 and the grammar description 26 is obtained using the acoustic segment model corresponding to this estimated section, and the recognition candidate with the highest likelihood is output as a recognition result. .

As described above, according to the present invention, it is possible to provide a method of creating an acoustic segment model in consideration of the relationship with the preceding and following segments and using the model to recognize the segment.

[0012]

As described above in detail, according to the present invention, in a technique for recognizing speech based on the trajectory of an acoustic segment, modeling is performed in consideration of the relevance of the acoustic features of the preceding and following segments. As a result, there is an advantage that it is possible to provide more excellent recognition performance in speech recognition using the same than an acoustic model represented by a conventional HMM.

An embodiment will be described below. Fifteen men and fifteen women were used for learning, and five men and five women were used for testing. For a 25 ms window of audio, the vector of 13 mel-opened cepstrum coefficients is 1
Calculated every 0 ms. In some experiments, so-called delta and acceleration factors were added to this static factor.
After parameterizing the words to emphasize speaker changes, the average vector was determined and the average vector was subtracted from the parameter vector for each frame. In this experiment, all models are context-dependent (triphones), each model is three-mixed, the HMM model has three states, and the segment model has the same number of parameters for the HMM model and the segment model. The HMM used a unique exponential interval model, and the segment model used a Galassian interval model. The segment model considered only the last 30 ms of the previous segment. This value was chosen to include the entire transition region, but avoided using distant acoustic data. The error rate when the phoneme model HMM is used is 15.47 for static parameters.
% And the static + で は + 静 的 parameter are 13.57%, and the error rates when the polynomial segment model is used are 11.53% and 10.18%, respectively, and when the model of the present invention is used. Are 10.05%,
It was 9.31%. It can be understood that the use of the segment model has an error rate of 25% or more than the use of the HMM model, and that the present invention further improves the error rate by 9 to 13%.

[Brief description of the drawings]

FIG. 1 is a diagram showing a range in which a locus of a characteristic parameter is obtained in an acoustic model used in the present invention.

FIG. 2 is a block diagram showing a process of generating an acoustic model used in the present invention.

FIG. 3 is a block diagram showing a functional configuration of a speech recognition device according to an embodiment of the present invention.

Claims (4)

[Claims] 1. A speech recognition apparatus for analyzing an inputted speech signal into speech acoustic feature parameters, and comparing the locus of the parameters with a model for each segment in units of phonemes, sub-words or words to perform recognition. In the first, including the end of the segment immediately before the segment
, A second section including the beginning of the segment immediately after the segment, or a third section including the end of the segment immediately before the segment and the beginning of the segment immediately after the segment.
A memory for storing a segment model representing the trajectory of the characteristic parameter for each segment label for at least one of the sections, a means for calculating the sound and audio parameters of the input sound signal, Means for calculating a trajectory for each section corresponding to the segment model in the memory; means for obtaining the likelihood of the calculated trajectory for each segment model in the memory; and recognition candidates using the likelihood A voice recognition device. 2. The speech recognition apparatus according to claim 1, wherein each of the segment models stored in the memory includes at least one of a label of each segment model, a label of a segment immediately before the segment model, and a label of a segment immediately after the segment model. A speech recognition apparatus characterized in that in the means for calculating the likelihood, at least one of the label information of the segments immediately before and after the corresponding segment is also considered. 3. A speech recognition method for analyzing speech acoustic feature parameters of an input speech signal and performing speech recognition for each segment based on phonemes, sub-words, or words based on the locus of the parameters. A first section including the end of the segment immediately preceding the segment, a second section including the beginning of the segment immediately following the segment, a third section including the end of the segment immediately preceding the segment and the beginning of the segment immediately following the segment.
A segment model representing the trajectory of the characteristic parameter is created in advance for at least one section of each section for each label of each segment, and this is stored in a memo. At the time of speech recognition, the speech acoustic parameters of the input speech signal are Calculating a trajectory of each of the sections corresponding to the segment model in the memory of the calculated voice acoustic parameter, obtaining a likelihood of the calculated trajectory for each segment model in the memory, A speech recognition method characterized by performing speech recognition using likelihood. 4. The speech recognition method according to claim 3, wherein the segment model is created by considering at least one of a label of the model, a label of a segment immediately before the model, and a segment label immediately after the model. A speech recognition method characterized in that in calculating a degree, likelihood calculation is performed in consideration of label information of segments immediately before and after a corresponding segment according to the model.