韵律增强型汉语语音合成系统

Mandarin text⁃to⁃speech system with prosody enhancement

端到端语音合成(TTS)系统可以直接根据给定的字素或音素序列生成语音。当前主流的端到端语音合成系统可以为英语生成近似于人类声音的语音。然而,中文的文本不同于这类以罗马字母为基础的语言(如英语),直接将端到端语音合成框架应用于汉语时,合成音频存在较为严重的韵律问题,如断句或停顿不恰当、自然度差等。为此,结合汉语的语言特性和韵律特性,提出一种神经网络端到端韵律增强型汉语语音合成系统,该系统使用从预训练Bert模型中提取的多层次上下文特征增强端到端汉语语音合成系统的输入。在汉语语音合成公开数据集上的实验结果表明,与当前主流的端到端语音合成系统相比,该韵律增强型汉语语音合成系统可以生成更加自然且富有表现力的语音。

The end⁃to⁃end text⁃to⁃speech(TTS)system can generate speech according to a given sequence of graphemes or phonemes.At present,the main current end⁃to⁃end TTS system can generate the speech that sounds akin to human voice for the English.However,the text of the Chinese is different from that of roman⁃letter based languages like the English.When the end⁃to⁃end TTS architecture is applied to mandarin speech synthesis,there are relatively serious prosodic problems such as inappropriate pauses and poor naturalness.That′s why a neural end⁃to⁃end mandarin TTS system with prosody enhancement is proposed in combination with the language and prosody features,which uses multi⁃level context features extracted from the pre⁃trained language model to enhance the input of the end⁃to⁃end mandarin TTS system.The results of the experiments conducted on a public Chinese speech synthesis dataset show that the system can generate more natural and more expressive mandarin speech in comparison with the state⁃of⁃the⁃art speech synthesis systems.