Word acoustic feature system and training method and system of word acoustic feature system

Abstract

The embodiment of the invention provides a training method of a word acoustic feature system. The method comprises the following steps: splicing word acoustic features output by a word acoustic feature system with a phoneme feature sequence output by a phoneme encoder to obtain a phoneme feature sequence with the word acoustic features, and splicing the phoneme feature sequence with actual rhythm features to obtain a phoneme feature sequence with rhythm and word acoustic features; adjusting the coding length, adding pitch and energy characteristics, and performing decoding to obtain a predicted Mel spectrum; and training the word acoustic feature system based on the actual Mel spectrum and the predicted Mel spectrum. The embodiment of the invention further provides a word acoustic feature system and a training system of the word acoustic feature system. According to the embodiment of the invention, the word acoustic features with word meanings and pronunciation are obtained by utilizing the trained word acoustic feature system, and the word acoustic features are more accurate by continuously training the word acoustic feature system, so that the speech synthesis quality is further improved during speech synthesis.

Description

technical field [0001] The invention relates to the field of intelligent speech, in particular to a word acoustic feature system, a training method and a system for the word acoustic feature system. Background technique [0002] End-to-end text-to-speech synthesis models with sequence-to-sequence architectures have achieved great success in generating natural speech. Word features are characterized by text analysis or word vector representations extracted from pre-trained models, and word vector encoders are then aligned and concatenated with phoneme feature sequences (output of phoneme encoders). Ways to obtain these eigenvectors include: [0003] Obtain word features through statistical methods, such as word frequency, etc., and then use text analysis methods to generate word feature vectors; [0004] Extract encoder output from common machine learning tasks (such as translation tasks) as word vectors; [0005] Use the BERT encoding layer to extract word vectors; [00...

AI Technical Summary

Problems solved by technology

[0009] In order to at least solve the problem that the existing models in the existing methods only focus on the meaning of the word and ignore the pronunciation of the word, making the feature vector less effective in improving the quality of speech synthesis

Images