Method for outputting phoneme probability by speech deep neural network model

Abstract

The invention discloses a method for outputting a phoneme probability by a speech deep neural network model. In an existing low-signal-to-noise-ratio actual use scene, the credibility of the probability of outputting phonemes of a voice deep neural network model is not high. The method comprises the following steps: firstly, adding confidence coefficient information to each original phoneme according to phoneme categories, and then comparing the confidence coefficient information with a set threshold value: when the confidence coefficient information is greater than or equal to the threshold value, keeping the probability corresponding to the original phoneme unchanged, and when the confidence coefficient information is lessthan the threshold value, adding a correction value on the basis of the probability of the original phoneme before correction to serve as an output phoneme probability; and finally, enabling the deep neural network to output a phoneme probability as a decoding basis. By adopting the method, after phoneme confidence correction, the voice probability of consonants in the phoneme probability correction is improved, the character recognition capability is improved, the phoneme confidence correction maintains the voice probability of vowels, and the occurrence of false activation is reduced.

Description

technical field [0001] The invention belongs to the technical field of computing, in particular to the technical field of speech deep neural network processing, and relates to a method for outputting phoneme probability by a speech deep neural network model. [0002] technical background [0003] Deep Neural Networks (DNNs) have been widely used in speech processing. The input of the speech DNN is the speech feature, and the output of the DNN is the probability of the phoneme. DNN outputs the probability of all phonemes once in each time interval, and the decoding algorithm decodes according to the output phoneme probability. The decoding method is a method of converting the probabilities of phonemes into words. Currently commonly used decoding algorithms are Beamsearch (beam search) and CTC (connection time series classifier) ​​and so on. Beamsearch is a heuristic graph search algorithm. In order to reduce the space and time used for searching, some nodes with lower proba...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to provide a method for outputting the phoneme probability of a voice deep neural network model in view of the disadvantage that the reliability of the probability of the output phoneme of the DNN is not high in the actual use scene of the existing low signal-to-noise ratio