Sound selection method and device for waveform splicing speech synthesis

Abstract

The invention provides a voice selection method and a device used for waveform splicing of voice synthesis. The voice selection method comprises steps of obtaining annotated information which is obtained via front end processing for a to-be-synthesized text; obtaining a pre-generated machine learning model; and according to the annotated information and the robot learning model, carrying out machine learning pre-selection so as to obtain sub-waveform fragments of a candidate voice. Thus, pre-selection effects during voice synthesis can be improved.

Description

technical field [0001] The invention relates to the technical field of speech synthesis, in particular to a sound selection method and device for waveform splicing speech synthesis. Background technique [0002] Speech synthesis, also known as text-to-speech (Text to Speech) technology, solves the main problem of how to convert text information into audible sound information. [0003] In speech synthesis, it is necessary to perform front-end processing on the input text first, then predict the acoustic parameters to obtain the acoustic parameters, and finally use the acoustic parameters to directly synthesize the sound through the vocoder, or select units from the sound library for waveform splicing. Compared with the voice synthesized by the vocoder, the synthesized voice based on waveform splicing has higher sound quality and better maintains the style of the original speaker. [0004] In the process of constructing a speech synthesis system based on waveform splicing, in...

AI Technical Summary

Problems solved by technology

[0006] (1) Each pre-selection process is independent of each other, and the information is not integrated and fully considered, so it is difficult to obtain a good pre-selection effect;

[0007] (2) The above-mentioned pre-selection process needs to adjust the threshold and weight, and the work of adjusting the threshold and weight requires a lot of meticulous manual work, and it is easy to lose sight of the other. After adjusting the threshold and weight for a sound bank, it is often necessary to readjust these parameter;

[0008] (3) Multi-step pre-selection is required, and the amount of calculation is large (especially KLD pre-selection);

[0009] (4) The engineering implementation of this method is relatively cumbersome, involving the maintenance of a large number of parameters, the code complexity is high, and it is difficult to maintain

Images