HMM与神经网络相融合的低资源语音合成方法  

Low Resource Speech Synthesis Method Based on HMM and Neural Network

作　　者：帕丽旦·木合塔尔 吾守尔·斯拉木[2] 买买提阿依甫 Palidan·Muhetaer;Silamu·Wushouer;Maimaitiayifu(School of Information Management,Xinjiang University of Finance&Economics,Urumqi Xinjiang 830012,China;College of Information Science and Engineering,Xinjiang University,Urumqi Xinjiang 830046,China)

机构地区：[1]新疆财经大学信息管理学院,新疆乌鲁木齐830012 [2]新疆大学信息科学与工程学院,新疆乌鲁木齐830046

出　　处：《计算机仿真》2021年第12期203-211,共9页Computer Simulation

基　　金：自治区天池博士计划(40050095)。

摘　　要：为了提高语音合成自然度和稳定性,提出HMM与深度神经网络相融合的,以维吾尔语作为实验语言的语音合成方法。基于深度学习的端到端语音合成方法存在生成速度慢、稳定性及可控性不够好,但是合成语音自然度高,而基于HMM的方法系统稳定性好,合成语音自然度不如端到端的方法。因此,系统前端部分利用HMM(马尔科夫模型)获取维吾尔语固有的语言特征,后端合成部分利用深度神经网络框架建立自回归模型。前端文本分析用HMM模型获取语言特征,后端合成用不同的神经网路模型,并进行了对比试验。最后对于实验结果进行了评测。实验结果验证了基于HMM+BiLSTM的语音合成方法的效果最好。In order to improve the naturalness and stability of speech synthesis, this paper proposes a speech syn-thesis method that combines HMM with a deep neural network and takes Uyghur as the experimental language. Theend-to-end speech synthesis method based on deep learning has the disadvantages of slow generation speed, poorstability, and controllability, but the naturalness of the synthesized speech is high, while the method based on HMMhas good system stability, and the naturalness of the synthesized speech is not as good as the end-to-end method.Therefore, the front-end part of the system uses HMM(Markov model) to obtain the inherent language features ofUyghur, and the back-end synthesis part uses the deep neural network framework to establish the autoregressive mod-el. HMM model is used in front-end text analysis to obtain language features, and different neural network models areused in back-end synthesis. Finally, the experimental results are evaluated. The experimental results verify that thespeech synthesis method based on HMM+BiL STM has the best effect.

关 键 词：神经网络 语音合成 低资源 语言特征 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]