一种基于WaveNet的藏语语音合成方法  

A Tibetan Speech Synthesis Method Based on WaveNet

作　　者：丁云涛 才让卓玛[2,3] 贡保加 才智杰 DING Yun-tao;CAI Rangzhuoma;GONG Baojia;CAI Zhijie(College of Computer Science and Technology,Qinghai Normal University,Xining Qinghai 810016,China;School of Computer Science and Engineering,Southwest Minzu University,Chengdu Sichuan 610041,China;The State Key Laboratory of Tibetan Intelligent Information Processing and Application,Xining Qinghai 810008,China)

机构地区：[1]青海师范大学计算机学院,青海西宁810016 [2]西南民族大学计算机科学与工程学院,四川成都610041 [3]藏语智能信息处理及应用国家重点实验室,青海西宁810008

出　　处：《计算机仿真》2023年第1期295-299,538,共6页Computer Simulation

基　　金：国家自然科学基金资助项目(61966031,61866032);青海省科技项目(2019-SF-129);藏文信息处理与机器翻译重点实验室(2020-ZJ-Y05)。

摘　　要：针对Griffin-Lim声码器合成语音保真度低、人工合成痕迹明显的问题,提出了一种基于WaveNet神经网络声码器的藏语语音合成方法。方法基于藏文文本特征及藏语Mel频谱,运用WaveNet的相位恢复能力实现了藏语的语音合成。先将藏文文本经字符编码转换为相应词向量,并通过卷积运算和注意力权重赋值完成文本特征提取;利用非线性变换和线性投影预测Mel频谱;利用WaveNet声码器合成藏语语音波形。实验数据表明,相对于Griffin-Lim算法合成的语音,所提方法所合成的语音效果更好。To solve the problem of low fidelity and obvious synthetic traces of speech synthesis with Griffin-Lim vocoder, this paper proposes a Tibetan speech synthesis method based on the neural network vocoder WaveNet. This method combines the phase information recovery capability of WaveNet. The characteristics of Tibetan sequence and Mel spectrum of Tibetan to synthesize the Tibetan speech. Firstly, the input sequence is converted into the corresponding word vector by character embedding, and the sequence feature extraction is completed by convolution operation and attention weight assignment. Secondly, the Mel spectrum is predicted by nonlinear transformation and linear projection. Finally, the WaveNet vocoder is used to synthesize the Tibetan voice waveform. The experimental data show that compared with the speech synthesized by Griffin-Lim algorithm, the speech synthesized by this method can get better results.

关 键 词：藏语语音合成 梅尔频谱 神经网络声码器 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]