JPH07104779A - Voice recognizing method - Google Patents

Abstract

PURPOSE:To prevent a deficiency in learning and to reduce the retrieval processing volume at the time of voice recognition by structuring a subword model which makes a vague acoustic phenomenon whose KANA(Japanese syllabary) notation can not be made to correspond to a subword nonambiguously correspond nonambiguously. CONSTITUTION:In a voice recognition device performs a voice recognizing process operates on the basis of voice recognition environment definitions 1 to perform voice recognition. In the voice recognition environment definitions 1, subword system definitions 12 have respective subwords and information for making subword labels and acoustic phenomena correspond to each other, and in the subword system definitions, subword labels which make acoustic phenomena whose KANA notations can not be made to corresponded to the subwords nonambiguously like 'souchi', 'sohchi'. etc., correspond nonambiguously are defined. Then hidden Markov models corresponding to the subword labels are made to learn and those subword labels are used to perform voice recognition wherein the vagueness of vocalization is absorbed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method using a hidden Markov model (hereinafter referred to as HMM).

[0002]

2. Description of the Related Art Conventionally, HMM (Hidden Markov Mode)
In the speech recognition method using l), a unit smaller than a word (hereinafter referred to as a subword), such as a phoneme, is used as a unit of an HMM that represents an acoustic phenomenon caused by a voice.
By using, it is possible to perform voice recognition of any vocabulary.

However, the voice data corresponding to a subword can be obtained only by cutting out a voice obtained by uttering a word or a sentence containing the subword (hereinafter, this work is referred to as labeling), and many voice data are obtained. Speech recognition based on HMM, which requires statistical learning using speech data, has a problem that the labeling work requires a lot of time and labor.

As a method for reducing this problem, there is connection learning. The connection learning uses that the HMM corresponding to the sentence or word is formed by connecting the HMM of the subword corresponding to the learning voice data (hereinafter referred to as the subword HMM),
It is possible to perform learning on each subword HMM by collectively learning HMMs corresponding to a sentence or a word including a plurality of subwords without giving information about an accurate existence section of a subword label that is a name of a subword. It can be done.

Therefore, in the speech recognition method using the above-mentioned connection learning, the labeling work for checking the existence section of the subword label can be omitted, and the learning using a large amount of speech data can be performed relatively easily. In the above speech recognition method, the learned speech data is converted into a subword label string corresponding to the data through a predetermined subword series and used for connection learning.

[0006]

By the way, in the above-mentioned conventional speech recognition method, when the subword label string given to the connection learning is generated from the kana notation, the kana notation and the acoustic phenomenon which is the actual uttered sound are complicated. However, learning may be insufficient. For example, as shown in FIG. 1, the part represented by “sachi” in kana notation is “s, o, u, c
It is converted into a subword label string "h, i", but it may be uttered as "so chi" or "so chi" depending on the habit of the speaker.

[0007] In addition, ice (koori) is "k, o, o, r,
Although it is converted to "i", there is a possibility that "kouri" or "kori" is actually uttered. As described above, there is an ambiguity that the subword cannot be uniquely determined only by the kana notation when converting the kana notation into the subword label string. In the conventional speech recognition method, as shown in FIG. 2A, the acoustic phenomenon and the subword label are associated with each other in a one-to-one manner. Therefore, if the kana notation and the acoustic phenomenon are complicated, a similar acoustic phenomenon may occur. , Different subword H
Assigned to MM. That is, there is a problem that learning for the same voice data may be distributed to a plurality of HMMs due to ambiguity in pronunciation, which may cause insufficient learning.

Further, as described above, if there is ambiguity that there are a plurality of subword labels corresponding to the same voice data, when performing voice recognition in consideration of this ambiguity, a plurality of subwords are recognized at the time of voice recognition. There is also a problem that it is necessary to search all the labels, which increases the search processing. The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a voice recognition method capable of preventing learning shortage and reducing the amount of search processing at the time of voice recognition.

[0009]

A speech recognition method according to the present invention is a speech recognition method in which a subword smaller than a word is set as a unit for expressing an acoustic phenomenon by a hidden Markov model. It is characterized by constructing a subword model that uniquely corresponds to an acoustic phenomenon that cannot be uniquely associated with the subword.

[0010]

According to the above method, a subword model is constructed in which an ambiguous acoustic phenomenon that cannot uniquely correspond to a kana description of an acoustic phenomenon and a subword can be associated. Therefore, all ambiguous acoustic phenomena can be learned by one model, and the audio data for learning will not be distributed to a plurality of subword models. In addition, the number of subword models to be searched for at the time of speech recognition absorbing the ambiguity of pronunciation is reduced. That is, insufficient learning is prevented, and the search processing amount at the time of voice recognition is reduced.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 3 is a diagram showing a functional configuration of a voice recognition device to which a voice recognition method according to an embodiment of the present invention is applied. In the speech recognition apparatus shown in this figure, 1 is a speech recognition environment definition having various definitions necessary for performing speech recognition processing, and 2 is speech recognition processing for performing speech recognition using an HMM in units of subwords. The voice recognition device according to the present embodiment is configured so that the voice recognition process 2 operates based on the voice recognition environment definition 1 to perform the voice recognition process.

In the voice recognition environment definition 1, 11 is a feature vector definition, and an analysis method for obtaining a feature vector including linguistic features of voice (for example, LPC (Li
nearPredictive Coding) information for selecting a cepstrum), and information regarding dimensions of various parameters obtained by the selected analysis method. Reference numeral 12 is a sub-word system definition, which has information for associating each sub-word label and the sub-word label with an acoustic phenomenon.

The above-mentioned information for associating is, for example, as shown in FIG. 2B, in kana notation "Ou" and "Oo".
"LOUL" is made to correspond to the subword label corresponding to, and the definition information shown in the following example makes it possible to express the part where ambiguity remains in kana notation with one subword label.

LgL: / g /, its nasalization, their intermediate vocalization LgyL: / gy /, its nasalization, their intermediate vocalization LwoL: / o /, / wo /, their intermediate Vocalizations LOUL: / o // u /, temporal fusion of / o / and / u /, / oo /, interim vocalizations LEIL: / e // i /, / e / and / I / temporal fusion, / ee /, their intermediate vocalizations LIUL: / i // u /, / i / and / u / temporal fusion, / y // uu /, those Intermediate vocalization LXIL: / i /, its unvoiced, those intermediate vocalizations LXUL: / u /, their unvoiced, their intermediate vocalizations LPL: Presence or absence of silent intervals

In the above definition information example, for example, the subword label "LOUL" is the generated sound / o // u /, / o.
It means to correspond to u /, / oo /, etc. Further, in the speech recognition environment definition 1 of FIG. 3, reference numeral 13 is an HMM structure definition, which has information on the number of states of the HMM and a method of expressing the output probability density distribution.

Next, in the voice recognition process 2, 15 is
An analysis process of converting an input voice (hereinafter, referred to as an input voice) into a characteristic parameter, 16 gives an HMM parameter (described later) by giving the characteristic parameter and linguistic subword label information corresponding to the parameter. This is a learning process for estimating. Reference numeral 17 is a recognition process, which recognizes the input voice by the HMM parameter estimated by the learning process 16 and the feature parameter supplied by the analysis process 15.

In the analysis process 15, a reference numeral 21 designates an AD (Analog) for band-limiting the input voice and converting it into digital data.
The to digital) conversion unit 22 is a characteristic parameter calculation unit that calculates a characteristic parameter corresponding to the input voice from the digital data output from the AD conversion unit 21.

In the learning process 16, 33 is label data having the subword label defined by the subword system definition 12 and the appearance time of the label in the input voice. An initial learning unit 23 performs initial learning based on the label data 33 and the characteristic parameters supplied from the characteristic parameter calculation unit 22 and outputs HMM parameters according to the learning.

Reference numeral 26 denotes learning kana notation character string data 32, which is a kana notation of the input voice, and learning label string data 34.
It is a kana notation label string conversion unit for converting into. Reference numeral 24 denotes a connection learning unit, which performs connection learning according to the HMM parameters supplied from the initial learning unit 23, the feature parameters supplied from the feature parameter calculation unit 22, and the learning label string data 34, and outputs the corresponding HMM parameters. Output. This HMM parameter represents the state transition probability of the HMM estimated by the learning process 16, the output density distribution for each state, etc., and is stored in the HMM parameter data 31.

Further, in the recognition processing 17, 27 is a kana notation label string conversion unit, and the subword system definition 12
Based on the above, the recognition kana notation character string data 35 representing the “acceptable grammar” used when performing the voice recognition in consideration of the ambiguity of utterance is converted into the recognition label string data 36. Further, 25 is the recognition label string data 36, the characteristic parameter calculated by the characteristic parameter calculation unit 22, and H
A recognition processing unit that performs voice recognition processing based on the MM parameter data 31 and outputs a recognition result.

In such a configuration, the analysis processing 15 will be described first. The input voice is band-limited by the AD converter 21 and converted into digital data. This digital data is supplied to the characteristic parameter calculation unit 22, and is subjected to an analysis process based on the analysis method and parameter dimensions defined by the characteristic vector definition 11. Then, the characteristic parameter corresponding to the input voice is calculated. This characteristic parameter is supplied to the learning process 16 and the recognition process 17.

In the learning process 16, the characteristic parameters are supplied to the initial learning section 23 and the connection learning section 24. In the initial learning unit 23, the feature parameter and the label data 3
From 3, the initial learning of the HMM for each subword label is performed. In the initial learning unit 23, Segmental k-means traininng procedure and Fo are used as learning algorithms.
The rward-Backward algorithm is used. These details can be found in LRRabiner, JGWilpon, and BHJuang, “A
segumental k-means training procedure for connected
word recognition ”(AT & T Technical Journal: vol.65
pp.21-31, (1986)), and Seiichi Nakagawa, "Speech recognition by probabilistic model" (IEICE, (1988)). As described above, in the initial learning unit 23, initial learning is performed and HMM parameters are obtained.

This HMM parameter is used by the connection learning unit 24.
And is used for connection learning together with the feature parameter and the learning label string data 34. Details of connected learning are described in, for example, Minami, Matsuoka, Kano, “Evaluation of connected learning HMM using unspecified speaker continuous speech database” (IEICE Technical Report, SP91-113, (1992)). There is. The connection learning unit 24 re-estimates the HMM parameters obtained by the initial learning and obtains the HMM parameter data 31.

On the other hand, in the recognition processing section 25 of the recognition processing 17, the feature parameter calculation section 22 supplies it based on the recognition label string data 36 in which "allowable grammar" is described and the HMM parameter data 31. The recognition processing of the sub-word label string corresponding to the feature parameter is performed. Details of the Viterbi algorithm used in this recognition processing are described in, for example, Seiichi Nakagawa, “Speech Recognition by Stochastic Model” (IEICE, (1988)).

As described above, since the subword label which absorbs the ambiguity of the utterance is defined by the subword system definition 12, it is possible to easily convert the kana notation into the subword label string. Further, since all the ambiguous acoustic phenomena can be represented by one subword label, a large amount of learning is performed without the voice data corresponding to one subword label being dispersed in a plurality of HMMs in the learning process. be able to.

Further, since all the ambiguous acoustic phenomena can be represented by one sub-word label, in the case of recognizing the ambiguity of utterance at the time of speech recognition, as in the conventional case, for example, a generated sound is generated. Since it is not necessary to allow both the subword label corresponding to / ou / and the subword label corresponding to utterance / oo /, it is possible to reduce the amount of calculation related to the search process.

[0027]

As described above, according to the present invention,
A subword model is constructed in which an ambiguous acoustic phenomenon that cannot be uniquely associated with a kana description of an acoustic phenomenon and a subword is uniquely associated. Therefore, all ambiguous acoustic phenomena can be learned by one model, and the audio data for learning will not be distributed to a plurality of subword models. In addition, the number of subword models to be searched for at the time of speech recognition absorbing the ambiguity of pronunciation is reduced.
Therefore, it is possible to prevent insufficient learning and reduce the amount of search processing during voice recognition.

[Brief description of drawings]

FIG. 1 is a diagram showing an example of conversion from kana notation to a subword label string.

FIG. 2 is a diagram showing a correspondence relationship between kana notation, an acoustic phenomenon, and a subword label.

FIG. 3 is a diagram showing a functional configuration of a voice recognition device to which a voice recognition method according to an embodiment of the present invention is applied.

[Explanation of symbols]

 1 voice recognition environment definition 2 voice recognition processing 11 feature vector definition 12 subword system definition 13 HMM structure definition 15 analysis processing 16 learning processing 17 recognition processing 21 AD conversion unit 22 feature parameter calculation unit 23 initial learning unit 24 connection learning unit 25 recognition processing Part 26, 27 Kana notation label string conversion unit 31 HMM parameter data 32 Learning kana notation character string data 33 Label data 34 Learning label string data 35 Recognition kana notation character string data 36 Recognition label string data

Claims (1)

[Claims] 1. In a speech recognition method in which a subword smaller than a word is set as a unit for expressing an acoustic phenomenon by a hidden Markov model, the kana notation of the acoustic phenomenon cannot be uniquely associated with the subword. A speech recognition method characterized by constructing a subword model that uniquely corresponds to an acoustic phenomenon.