Prosodic hierarchy labeling method and model training method and device

Abstract

The invention discloses a prosodic hierarchy labeling method. The method comprises the steps as follows: acquiring to-be-labeled text data and audio data which have a corresponding relation; extracting a to-be-labeled text characteristic set of each word according to the to-be-labeled text data; extracting an acoustic characteristic set of each word according to the audio data; acquiring a prosodic hierarchy structure by a prosodic hierarchy labeling model according to a word identification of each word, the to-be-labeled text characteristic set of each word and the acoustic characteristic setof each word. The invention also discloses a model training method, a prosodic hierarchy labeling device and a model training device. The prosodic hierarchy labeling model is established by combination of text characteristics and acoustic characteristics, richer characteristics can be provided for prosodic hierarchy labeling, the prosodic hierarchy labeling accuracy can be improved, and the voicesynthesis effect can be enhanced.

Description

technical field [0001] The present invention relates to the field of artificial intelligence, in particular to a prosody-level labeling method, a model training method and related devices. Background technique [0002] In order to realize a high-quality speech synthesis system, it is very important to accurately label a large amount of data with a prosodic hierarchical structure. The prosodic hierarchical structure is to model the rhythm of the speech and its pauses. A method that can accurately and automatically label the prosodic hierarchical structure is helpful for fast It is of great significance to build a speech synthesis corpus and improve the naturalness of speech synthesis. [0003] At present, the automatic labeling of the prosodic hierarchical structure needs to use machine learning methods to train an automatic labeling model. There are two main types of feature selection. One is to use text features, first segment words, and then extract text features of words,...

AI Technical Summary

Problems solved by technology

[0004] However, in practice, the labeling task only uses text data, and does not take into account the prolongation of the duration of the previous syllable at the boundary of the prosodic hierarchy and the phenomenon of short pauses at the boundary of intonation phrases, and it is difficult to accurately use acoustic features. At the same time, the three-level prosodic hierarchical structure is marked, and the internal relationship between text features and acoustic features is ignored, which reduces the labeling effect of the prosodic hierarchical structure and affects the quality of the corpus on which speech synthesis depends.

Images