Speech recognition optimization decoding method integrating guide probability

Abstract

The invention discloses a speech recognition decoding method integrating guide probability and provides a guide probability model for overcoming the insufficiency that the traditional speech recognition system insufficiently utilizes position information of a speech frame in an acoustic characteristic space, and for the purpose of describing the probability that the speech frame belongs to different parts of the acoustic characteristic space and guiding a decoding process. The method comprises the following steps of training a universal background model for describing the whole acoustic characteristic space, computing a main gaussian component of the speech frame in the universal background model, utilizing an acoustic model of a recognition system to forcibly segment a training corpus, obtaining a phoneme of the speech frame, counting response frequency of the phoneme and main gaussian, normalizing the respond frequency, obtaining the guide probability, fusing the guide probability to total score computation of a speech recognition path, and guiding a decoder to enhance or weaken the path.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to the fields of acoustic modeling and decoding for speech recognition. Background technique [0002] At present, speech recognition systems generally use Hidden Markov Model as the basic model of acoustic modeling and decoding. In order to consider the impact of contextual pronunciation on speech units, triphone models are often used to improve the system recognition rate. But after considering the context, the number of models and the size of the parameters increase dramatically. Taking the Chinese large vocabulary continuous speech recognition system as an example, the basic phoneme set only contains 191 initials and tonal finals, while the corresponding triphone models total more than 200,000. Even after the parameter sharing of the model layer, state layer and Gaussian component layer, the scale of parameters is still huge. This will not only bring about insufficient parame...

AI Technical Summary

Problems solved by technology

[0008] The purpose of the present invention is to solve the shortcomings of the lack of using the position information of the speech frame to be recognized in the acoustic feature space and the lack of enhanced search for some local spaces in the existing speech recognition decoding technology

Images