JPH0573094A - Continuous speech recognizing method - Google Patents

Abstract

PURPOSE:To improve the processing speed and document recognition rate by correcting a phoneme series signal by acoustic processing by a neural network by learning based upon a tutor signal. CONSTITUTION:The neural network obtains a phoneme series, corrected by the learning based upon the phoneme series signal and an actually voiced phoneme series signal, from an acoustic processing part 1 and supplies this corrected phoneme series signal as a phoneme series signal to a language processing part 2. The phoneme series is corrected while the phoneme series signal is learnt by using the speech signal that a speaker actually voices as the tutor signal. Therefore, this learning is improved in correcting ability as the process is repeated more. Consequently, word candidates at the language processing part 2 are decreased to make the processing speed fast and also improve the document recognition rate at the same time.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a continuous speech recognition method, and more particularly to a conversational speech recognition method.

[0002]

2. Description of the Related Art Among the speech recognition methods, a continuous speech recognition method for recognizing a speech in which a word is uttered in succession is a continuous word speech recognition for recognizing a relatively small number of vocabulary words, and further for adding linguistic knowledge There is speech recognition that recognizes the meaning and content of a relatively large number of words. FIG. 2 shows the configuration of the conversational voice recognition system. This configuration shows a hierarchical model in which the acoustic processing unit 1 and the language processing unit 2 perform distributed processing in hierarchical layers. The sound processing unit 1 is
Feature extraction is performed from the input voice, and the voice signal is segmented and converted into a phoneme string by phoneme recognition. The language processing unit 2 creates a word or word string candidate from a dictionary and phonological rules, corrects the phoneme string by using linguistic information such as syntax, meaning, and context, and outputs it as a sentence.

In addition to this hierarchical model, there are a blackboard model that uses a database common to acoustic levels and language levels, and a network model that uses a network at language levels.

[0004]

SUMMARY OF THE INVENTION In the conventional method,
It is difficult for the acoustic processing unit 1 to perform complete phoneme recognition,
The output phoneme sequence contains an error. Therefore, since the language processing unit processes on the assumption that an error is included, the word that does not completely match is also added as a candidate, and the incorrect candidate is removed from the language information such as the syntax and the meaning.

For this reason, when the phoneme recognition rate in the sound processing section is lowered, the number of word candidates to be raised in the language processing section is increased, and there is a problem that the processing speed is lowered and the sentence recognition rate is lowered.

It is an object of the present invention to provide a continuous speech recognition method with improved processing speed and sentence recognition rate.

[0007]

When a phoneme string signal is extracted by acoustic processing on an input speech signal and sentence recognition is performed by language processing on this phoneme string signal, the actual speech signal of a speaker is learned as a teacher signal to learn the above. It is characterized in that language processing is performed with the phoneme string signal modified by a neural network that modifies the phoneme string signal.

[0008]

Operation: A phoneme string signal corrected by supervised learning by a neural network is obtained from a phoneme string signal that has been acoustically processed, and this is used as a phoneme string signal for language processing. To reduce word candidates in the language processing unit.

[0009]

1 is a block diagram showing an embodiment of the present invention. In the figure, the sound processing unit 1 and the language processing unit 2 perform the same processing as the conventional hierarchical model.

The neural network 3 is a sound processing unit 1.
A phoneme string corrected by learning is obtained from the phoneme string signal and the phoneme string signal actually uttered, and the corrected phoneme string signal is given to the language processing unit 2 as a phoneme string signal.

The phoneme string correction by the neural network 3 is performed by supervised learning, and a corrected phoneme string signal for a phoneme string acoustically processed by a speaker using a voice signal actually generated as a teacher signal is corrected by learning. Alternatively, the learning is corrected by using the error from the actual voice as the teacher signal.

By interposing the neural network 3 as described above, the tendency of the error to the phoneme string generation or the error itself generated in the acoustic processing unit 1 is learned by the neural network, and this learning is stepwise as the processing is repeated. The ability to correct is enhanced and corrects phoneme string errors. As a result, the number of word candidates in the language processing unit 2 is reduced, the processing speed is increased, and the sentence recognition rate is also increased.

The present invention is not limited to the hierarchical model, and can be applied to the sound level processing in the blackboard model or the network model to obtain the same effect.

[0014]

As described above, according to the present invention, a phoneme string signal obtained by acoustic processing is corrected by supervised learning by a neural network, and the corrected phoneme string signal is subjected to language processing. The number of word candidates in the language processing unit can be reduced to improve the processing speed and the sentence recognition rate even in the case of processing error.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an embodiment of the present invention.

FIG. 2 is a block diagram showing a conventional method.

[Explanation of symbols]

1 ... Acoustic processing unit, 2 ... Language processing unit, 3 ... Neural network.

Claims (1)

[Claims] 1. A phoneme string signal is extracted by acoustic processing of an input speech signal, and sentence recognition is performed by language processing of the phoneme string signal. When the phoneme string signal is learned by learning an actual speech signal of a speaker as a teacher signal. A continuous speech recognition method characterized by performing language processing on a phoneme string signal modified by a modifying neural network.