Rhythmical pause information determination method and device

Abstract

The invention provides a rhythmical pause information determination method and a device. The rhythmical pause information determination method comprises the steps of extracting rhythmical prediction features of a to-be-synthesized text; selecting a self-adaptive rhythmical prediction model corresponding to a selected speaker; and inputting the rhythmical prediction features of the to-be-synthesized text into the self-adaptive rhythmical prediction model corresponding to the selected speaker so as to determine the rhythmical pause information of the to-be-synthesized text. According to the technical scheme of the invention, the problem that the rhythmical training data for an acoustic model and the rhythmical training data for a rhythmical model are inconsistent can be solved. Meanwhile, the rhythmical rhyme and the synthetic fluency are improved. Moreover, the self-adaptive rhythmical prediction model corresponding to the selected speaker is adopted, so that the synthetic effect in a multi-speaker handover situation is more natural.

Description

technical field [0001] The invention relates to the technical field of speech synthesis, in particular to a method and device for determining prosodic pause information. Background technique [0002] The purpose of speech synthesis is to convert text into speech and play it to users, and the goal is to achieve the effect of live text broadcasting. An important module in the speech synthesis link is to predict the prosodic pauses of the text to be synthesized, and then generate synthetic speech according to the predicted prosodic pauses. [0003] Currently, prosody prediction in speech synthesis is implemented based on statistical machine learning, and the process includes preparing training data, training prosody prediction models, and performing prosody predictions based on the trained models. [0004] However, in the prior art, the prosodic pause pattern trained in the prosody prediction model does not match the prosodic pause pattern trained in the acoustic model. The r...

AI Technical Summary

Problems solved by technology

This method solves the problem that the prosody training data used by the acoustic model and the prosody model are inconsistent, improves the prosody rhythm, improves the fluency of synthesis, and uses the speaker's own adaptive prosody prediction model, making it possible for multiple speakers to switch. Synthetic effect is more natural

Images