End-to-end voice synthesis method and system based on DNN-HMM bimodal alignment network

Abstract

The invention discloses an end-to-end voice synthesis method and system based on a DNN-HMM bimodal alignment network, and belongs to the field of intelligent voice interaction. According to the method, a frame length prediction module is used for replacing a traditional end-to-end attention autoregressive structure, and a convolutional change module and a bidirectional long-short-term memory network are used for constructing an encoder and a decoder, so that a large number of model parameters are reduced. On the basis that a phoneme frame length sequence is obtained through DNN-HMM bimodal alignment network training, an end-to-end voice synthesis model is trained, and thus the process that a traditional end-to-end voice synthesis model obtains text and audio alignment information in an autoregressive attention mode is avoided. The trained model not only can ensure the high naturalness of the audio synthesized by the end-to-end model, but also can greatly reduce the consumption of computing resources and the time proportion during voice synthesis, so that the end-to-end voice synthesis technology can be deployed on hardware with low computing resources.

Description

technical field [0001] The present invention relates to the field of intelligent voice interaction, further to the field of computer intelligent voice synthesis, and in particular to an end-to-end voice synthesis method and system based on a DNN-HMM dual-mode alignment network. Background technique [0002] In recent years, with the rise of deep learning, deep network models have dominated many fields of machine learning. Text to Speech (TTS), the process of synthesizing artificial speech from text symbols, is gradually being replaced by end-to-end deep neural networks. In the early days when people explored speech synthesis, scholars proposed speech synthesis methods based on statistical parameters. The speech synthesis method based on statistical parameters is mainly based on the parameter representation of speech features, such as Mel spectrum, fundamental frequency and other acoustic feature parameters, through hidden Markov model (HMM) modeling and related features of ...

AI Technical Summary

Problems solved by technology

[0007] In order to solve the problem that in the existing speech synthesis technology, due to the high complexity of the model, it cannot be used on low computing resources, and the effect of synthesizing long sentences is not ideal, the present invention proposes a dual-modal alignment based on DNN-HMM The end-to-end speech synthesis method and system of the network, on the basis of the DNN-HMM dual-modal alignment network training to obtain the phoneme frame length sequence, and then train the end-to-end speech synthesis model, thereby avoiding the traditional end-to-end speech synthesis model A process for informing text and audio alignment via autoregressive attention

Images