Method and system for detecting phonetic frequency spectrum wave crest and phonetic identification

Abstract

The invention provides a method and a device for detecting voice frequency spectrum wave crest as well as a voice recognition method and a system thereof. The method for detecting voice frequency spectrum wave crest comprises the following steps: detecting voice frequency spectrum wave crest candidates from the power spectrum of the voice; eliminating noise wave crests in the voice frequency spectrum wave crest candidates according to the spaces between the wave crests and adjacent wave crest positions; and detecting the voice frequency spectrum wave crest. In the invention, during the detection of the voice frequency spectrum wave crest, the limitations of the spaces between the wave crests and adjacent wave crest positions are utilized to remove the noise wave crests. Furthermore, the acquired power value of the voice frequency spectrum wave crest replaces the entire power spectrum, and is used for extracting the mel cepstrum coefficient characteristics of voice, so as to increase the noise-proof robustness of voice recognition under the circumstances that voice characteristic dimension is not increased.

Description

technical field [0001] The present invention relates to information processing technology, in particular, to the detection of the spectral peak of speech and the speech recognition technology using the spectral peak information of speech. Background technique [0002] The goal of Automatic Speech Recognition (ASR) technology is to enable computers to recognize continuous speech spoken by people. Usually, the automatic speech recognition process includes two stages of template generation and matching recognition. In the template generation stage, a template for comparison is established according to the spectral characteristics of the sample speech; in the recognition stage, when the speaker's speech is input into the computer, the computer's automatic speech recognition system extracts the features of these speech and uses This is compared with the pre-stored voice template to find the most matching and closest voice sample, so as to know the meaning of the input voice, and...

AI Technical Summary

Problems solved by technology

Some peaks in the noisy speech spectrum are caused by noise, and if the noise-induced peaks are mistaken for speech, it will lead to poor recognition performance

[0010] (2) The dimensionality of speech features cannot be increased too much

At present, most of the robust front-ends using spectral peak information combine the features of purely using spectral peak information with the traditional Mel scale cepstral coefficient, so the feature dimension will increase

Images