| 张维昭,李俊帜.完全端到端的藏语安多方言语声合成*[J].,2025,44(5):1251-1262 |
| 完全端到端的藏语安多方言语声合成* |
| Speech synthesis for Tibetan Amdo dialect base on complete end-to-end |
| 投稿时间：2024-05-21 修订日期：2025-08-28 |
| 中文摘要: |
| 目前藏语语声合成研究多面向卫藏方言，而对安多和康方言研究相对较少。在分析藏文特点的基础上，该文首先设计并构建一个大规模标准安多方言语声合成语料库(TACSS)，总时长为18.6 h。然后，设计了基于计算机可识别机读音标的SAMPA-AT和基于藏文构件的两种字素到音素(G2P)转写方案。最后，采用完全端到端语声合成模型VITS，实现了藏语安多方言的语声合成。与此同时，该文还比较了基于SAMPA-AT和基于藏文构件的两种G2P转写方案的优劣。实验结果表明，与两阶段语声合成模型相比，VITS在藏语安多方言语声合成任务上具有更好的表现。通过采用基于藏文构件的转写方案，该文提出的完全端到端藏语安多方言语声合成模型平均意见得分最优值为4.59。 |
| 英文摘要: |
| Current research on Tibetan speech synthesis is mostly based on the ü-Tsang dialect, while there are relatively few related studies on the Kham and Amdo dialects. Based on analyzing the characteristics of Tibetan, this paper first designs and constructs a large-scale standard Tibetan Amdo dialect corpus for speech synthesis (TACSS), with a total duration of 18.6 hours. Secondly, this paper designs two grapheme-to-phoneme (G2P) transcription schemes, one based on SAMPA-AT which is computer-recognizable machine-readable phonetic symbols, and another one based on Tibetan components. Finally, the complete end-to-end speech synthesis model VITS was used to realize speech synthesis of the Amdo dialect of Tibetan. At the same time, this paper also compares the advantages and disadvantages of two G2P transcription schemes which based on SAMPA-AT and Tibetan components. The experimental results show that compared with the two-stage speech synthesis model, VITS has better performance in the speech synthesis task of Tibetan Amdo dialect. By adopting a transcription scheme based on Tibetan components, the optimal mean opinion score (MOS) of the complete end-to-end Tibetan Amdo dialect speech synthesis model proposed in this paper is 4.59. |
| DOI：10.11684/j.issn.1000-310X.2025.05.016 |
| 中文关键词: 语音合成 藏语安多方言 端到端 语料库 藏文文本转写 |
| 英文关键词: Speech synthesis Tibetan Amdo dialect End-to-end Corpus Tibetan text transcription |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| 摘要点击次数: 420 |
| 全文下载次数: 207 |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
