A speech synthesis method and device

Abstract

The invention provides a voice synthesis method and a voice synthesis device. The method comprises the following steps: selecting candidate voice units for a voice to be synthesized from a voice base to form a candidate space through a first pre-trained model; and selecting the voice unit from the candidate space for splicing through a second pre-trained model, so that the search cost of a sequence consisting of the selected voice units is the best, wherein at least one of the first model and the second model is a neural network model. The voice synthesis method and the voice synthesis device can improve the naturalness and the expressive force of a finally synthesized voice.

Description

【Technical field】 [0001] The invention relates to the field of computer application technology, in particular to a speech synthesis method and device. 【Background technique】 [0002] With the advent of the mobile era, people's demand for speech synthesis is increasing, such as reading aloud novels, voice navigation, etc., all require speech synthesis. Moreover, people are not only satisfied with clarity and intelligibility for speech synthesis, but also require the synthesized speech to have better naturalness and expressiveness. [0003] For speech synthesis, it is first necessary to process the input text, including preprocessing, word segmentation, part-of-speech tagging, phonetic notation, prosodic level prediction, etc., then predict the acoustic features corresponding to each unit through the acoustic model, and finally use the acoustic parameters to pass the acoustic Speech is synthesized by the coder, or the appropriate speech unit is selected from the corpus for sp...

AI Technical Summary

Problems solved by technology

In the existing implementation, HMM (Hidden Markov Model, Hidden Markov Model) is used in the pre-selection process of the speech unit and the search process of the candidate space, but since the states in the HMM model are independent of each other, and it is based on The shallow modeling of the decision tree and the linear division of the feature space lead to low modeling accuracy in the case of complex text context features, resulting in smoother and poorer expressiveness in the final synthesized speech

Images