Sound selection method for waveform concatenation speech synthesis

Abstract

The invention discloses a sound selection method for waveform concatenation speech synthesis. The method comprises the following steps of: on the basis of an original audio, carrying out hidden markov based model training so as to obtain an acoustic model set and a corresponding characteristic decision tree; inputting a plurality of training texts and on the basis of the characteristic decision tree, searching to obtain related acoustic models so as to obtain corresponding target voice and target syllables; according to similarity of the target voice and corresponding candidate primitives and likelihood probability of each acoustic parameter of the candidate primitives under a current acoustic model, training to obtain a similarity classifier; inputting a random text to be synthesized, removing the dissimilar candidate primitives on the basis of the similarity classifier, selecting the optimal primitive from the residual candidate primitives by utilizing a concatenation cost minimization rule and carrying out concatenation to obtain synthetic speech. The adoption of the method disclosed by the invention can synthesize speech with higher tone quality.

Description

technical field [0001] The invention relates to the field of intelligent information processing, in particular to a sound selection method for waveform splicing speech synthesis. Background technique [0002] Speech is one of the main means for human beings to exchange information. Speech synthesis technology mainly enables computers to generate continuous speech with high definition and high naturalness. In the development of speech synthesis technology, the early research mainly used the parametric synthesis method, and later with the development of computer technology, the synthesis method of waveform splicing appeared. With the continuous increase of the corpus, the number of candidate primitives is also increasing. How to select the best primitives for splicing according to the input text has attracted more and more attention. [0003] The parametric speech synthesis system based on hidden Markov model and the splicing system based on primitive selection are the most m...

AI Technical Summary

Problems solved by technology

[0004] 1. The sound selection method does not reflect the perception of the human ear. Getting a high score in the existing sound selection method does not mean that a sound that is more suitable for human hearing has been selected;

[0005] 2. The sound selection method adopts the method of factor weighted superposition to select the sound, that is, the sub-costs are calculated for each feature of the primitive, and then given weights respectively, and then superimposed into a total sound selection cost to select the sound. This method assumes all factors The effect on the acceptance of primitives is linearly additive, which is clearly not true

Images