JPH0612090A - Speech learning system - Google Patents

Abstract

PURPOSE:To enable high-accuracy learning by invetigating whether learning data are suitable for learning a standard model or not for the unit of a phoneme. CONSTITUTION:At a feature analysis part 11, a speech signal S is converted to a feature vector time sequence V by using a mel cepstrum method. At a recognition part 13, the feature vector time sequence V and a standard model P in a standard model storage part 12 are inputted, the feature vector time sequence V is divided for the unit of music by observation, similarity gn (n=1, 2...N; N is the number of phonemes) is obtained by forward/backward algorithm described in a material 4 while using the standard model corresponding to the phoneme for feature vector time sequences Vn (n=1, 2...N) in respective phoneme blocks to output as a similarity sequence G={g1, g2...gn}. At a learning part 14, the feature vector time sequency V and the similarity sequence G are inputted and when each similarity gn is larger than a fixed value K decided in advance, learning is performed by using the feature vector time sequence V.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice learning system for learning a standard model used in voice recognition and the like.

[0002]

2. Description of the Related Art A large vocabulary speech recognition method that recognizes many vocabulary words by using a standard model created from learning data uttered in advance, such as phonemes as recognition units.
Units smaller than words are commonly used. Less than,
The term "phoneme" means not only the minimum basic unit of speech in the phonological sense, but also a wider range of units of speech including syllables and concatenation of plural phonemes. As a method of using a phoneme as a recognition unit, for example, Watanabe, Yoshida, Koga et al.
II Vol. J72-D-II No. 8 1989 8
The method described in the article “Large-vocabulary speech recognition using HMM in semisyllabic units” (hereinafter referred to as Document 1) published on page 1264-1269 of the moon is mentioned. In this method, a standard model is created for each semi-syllable (hereinafter referred to as a phoneme), which is a kind of phoneme, using a plurality of learning data uttered in word units. At the time of recognition, a word-by-phoneme dictionary is used to combine standard models to create a word-by-word model, and this word model is used to recognize unknown word speech. In this method, correctly uttered learning data is required to create a highly accurate standard model.

As a method for determining whether the uttered learning data is appropriate for learning the standard model, for example, Fujimoto,
Sato et al., "The Simplified Word Speech Recognition Method for Speech Recognition Boards", published on pages 85-86 of the Acoustical Society of Japan 1986 Spring Research Presentation Collection Ι, March 1986, pages 85-86. The method described in (1) is mentioned. This is because when a standard model has already been created for the words to be learned, the newly uttered learning data is recognized using the standard model, and the similarity between the standard model of the same word and the learning data (for example, , The sign of the distance is inverted) is equal to or greater than a certain threshold, the standard model is averaged, that is, the standard model is learned.

[0004]

In the method of the above-mentioned document 2, since it is determined whether or not the learning data is used for learning, it is determined in the learning data that the standard model and the learning data are similar to each other. Data for a particular phoneme
Even if it is inappropriate due to an erroneous utterance or the inclusion of noise, if the overall similarity is high, it may be used as learning data. This has a problem that the standard model for a specific phoneme is learned by inappropriate data in the method using a phoneme as a recognition unit as described in Document 1.

It is an object of the present invention to provide a speech learning method for performing highly accurate learning by checking whether or not the uttered learning data is suitable for learning the standard model in phoneme units.

[0006]

According to a first aspect of the present invention, there is provided a feature analysis unit for analyzing an input voice signal and outputting a feature vector time series as learning data, and a standard model created in advance in units of phonemes. A standard model storage unit for storing, a recognition unit that divides the learning data into units of the phonemes, obtains a degree of similarity of the divided learning data with respect to the standard model, and the learning data output from the recognition unit is configured. And a learning unit that performs learning using the learning data when the degree of similarity to the phoneme is larger than a predetermined value.

A second aspect of the present invention is a standard analysis unit that analyzes an input voice signal and outputs a feature vector time series as learning data, and a standard model that stores a phoneme standard model created in advance. A model storage unit, a recognition unit that divides the learning data into units of the phonemes, and obtains a similarity of the divided learning data to the standard model, and the similarity output from the recognition unit is a predetermined value. And a learning unit that performs learning by using data excluding portions corresponding to smaller phonemes.

[0008]

First, the first invention will be described. It is assumed that the HMM described in Reference 1 is used as a standard model of a phoneme unit used for recognizing spoken learning data, and that the learning method described in Reference 1 is used in advance. The HMM is a type of state transition network, and a state transition probability and a vector appearance probability are defined for each state.

The learning data is obtained by Furui's method in 1985 using the method by the mel cepstrum described on pages 154-160 of "Digital Speech Processing" (hereinafter referred to as Reference 3) published by Tokai University Press. It is converted into a sequence and further divided into phonemes. The phoneme-based division is performed by inspection, S. E. Levinson, L .; R.
Rabiner, and M.M. M. Sondi et al., T
he Bell System Technical
Journal, Vol. 62, No. 4, Apr. 1983, page 1035-1074, "An
Introduction to the Appli
Cation of the Theory of P
robabilistic Functions of
a Markov Process to Auto
"matic Speech Recognition"
(Hereinafter referred to as Literature 4) Viterbi (Vi
terbi) algorithm.
When using the Viterbi algorithm, the optimal state transition path in the model created by combining the standard models for the phonemes that make up the training data is found using the Viterbi algorithm, and the standard model for each phoneme unit on that path is obtained. It is possible to divide into phonemes by obtaining the times in the learning data corresponding to the beginning and end of the. After the division, the similarity of the data corresponding to the phoneme in the learning data to the standard model is obtained for each phoneme. When the learning data is divided using the Viterbi algorithm, the likelihood in the terminal state on the optimum state transition path of the standard model for each phoneme, which is already obtained at the time of division, can be used as the similarity. Alternatively, the forward-backward (forward-backwar) as described in Reference 4 is applied to the learning data of each divided phoneme section.
d) The degree of similarity can also be obtained using an algorithm. Learning is performed using the uttered learning data only when the degree of similarity to all or a specific standard model is larger than a predetermined constant value.

Next, the second invention will be described. In this case, when the phoneme-based similarity is smaller than a predetermined constant value, learning is performed using learning data excluding the voice section corresponding to the phoneme. As a result, learning is performed with high accuracy using only the correctly uttered learning data.

[0011]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the first invention. This embodiment includes a feature analysis unit 11, a standard model storage unit 12, a recognition unit 13, and a learning unit 14.

The standard model storage unit 12 holds a standard model P in advance. The learning voice signal S is input to the feature analysis unit 11. The feature analysis unit 11 converts the audio signal S into a feature vector time series V by using the method based on the mel cepstrum as described in Reference 3. In the recognition unit 13, the feature vector time series V and the standard model P in the standard model storage unit 12 are input, the feature vector time series V is divided into phoneme units by observation, and the feature vector time series V _n (V _n ( n = 1, 2, ..., N,
N is the number of phonemes), and using the standard model for that phoneme, the degree of similarity g _n (n = 1, 2, ...) By the forward backward algorithm described in Reference 4.
, N) is obtained, and the similarity sequence G = {g ₁ , g ₂ , ...
, G _N }. In the learning unit 14, the feature vector time series V and the similarity sequence G are input, and each similarity g _n
Is larger than a predetermined constant value K, learning is performed using the feature vector time series V.

FIG. 2 is a block diagram showing an embodiment of the second invention. This embodiment includes a feature analysis unit 11, a standard model storage unit 12, a recognition unit 13, and a learning unit 14a.

The standard model storage unit 12 holds a standard model P in advance. The learning voice signal S is input to the feature analysis unit 11. The feature analysis unit 11 converts the audio signal S into a feature vector time series V by using the method based on the mel cepstrum as described in Reference 3. In the recognition unit 13, the feature vector time series V and the standard model P in the standard model storage unit 12 are input, the feature vector time series V is divided into phoneme units by observation, and the feature vector time series V _n (V _n ( n = 1, 2, ..., N,
N is the number of phonemes), and using the standard model for that phoneme, the degree of similarity g _n (n = 1, 2, ...) By the forward backward algorithm described in Reference 4.
, N) is obtained, and the similarity sequence G = {g ₁ , g ₂ , ...
, G _N }. In the learning unit 14a, the feature vector time series V and the similarity sequence G are input, and the similarity g _n
Learning is performed using a feature vector time series obtained by removing the feature vector time series V _n for a phoneme smaller than a predetermined fixed value K from the feature vector time series V.

Although voice recognition has been described above, the present invention can also be used for selection of voice data that becomes a standard in learning a synthesis model used in voice synthesis.

[0017]

As described above, according to the present invention, it is possible to realize highly accurate learning of the standard model in the unit of phoneme by checking whether or not the uttered learning data is appropriate for the learning of the standard model. It has the effect of being able to.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a first invention.

FIG. 2 is a block diagram showing an embodiment of the second invention.

[Explanation of symbols]

 11 Feature Analysis Unit 12 Standard Model Storage Unit 13 Recognition Unit 14, 14a Learning Unit G Similarity Sequence P Standard Model S Learning Speech Signal V Feature Vector Time Series

Claims (2)

[Claims] 1. A feature analysis unit that analyzes an input voice signal and outputs a feature vector time series as learning data, a standard model storage unit that stores a standard model created in advance for each phoneme, and A recognition unit that divides the learning data into the units of the phonemes and obtains the similarity of the divided learning data to the standard model, and a similarity to the phonemes that constitute the learning data output from the recognition unit are predetermined. And a learning unit that performs learning using the learning data when the value is larger than a value. 2. A feature analysis unit which analyzes an input voice signal and outputs a feature vector time series as learning data, a standard model storage unit which stores a standard model in units of phonemes created in advance, The learning data is divided into the units of the phonemes, and the recognition unit obtains the similarity of the divided learning data with respect to the standard model, and the recognition unit outputs the similarity corresponding to a phoneme having a smaller value than a predetermined value. A speech learning method comprising: a learning unit that performs learning using data excluding a portion.