Zhou, Wangyiyao (2024) From Tolkien’s Novel to Synthetic Speech: Developing TTS Systems for Quenya. Master thesis, Voice Technology (VT).
|
PDF
MSc-5480698-W-ZHOU.pdf Download (2MB) | Preview |
Abstract
Text-to-speech (TTS) technology has been successfully implemented in various applications, serving as a means to preserve endangered languages and cultures. However, TTS systems for constructed languages have not been extensively studied. J.R.R. Tolkien created Quenya for the elves in his novels, inspired by the phonetic patterns and structures of Finnish. While enthusiasts have developed courses based on Tolkien's materials, attempts to synthesize Quenya speech remain limited. This study uses the articulatory features as inputs for speech synthesis and evaluates the outcomes of applying transfer learning from models based on more resourced languages. Using the IMS-Toucan system from the University of Stuttgart, based on the FastSpeech 2 architecture, this research developed a TTS system for Quenya by fine-tuning three models with a 34-minute Quenya dataset: one pre-trained on Finnish, one pre-trained on English, and a multilingual model. The results showed that the Finnish fine-tuned model produced better speech than the English model, while the multilingual model produced the most natural and accurate speech. This study provides insights for developing TTS systems for other constructed and ancient languages requiring phonetic reconstruction.
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Do, T.P. |
| Date Deposited: | 18 Jul 2024 19:34 |
| Last Modified: | 18 Jul 2024 19:34 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/531 |
Actions (login required)
 |
View Item |