Researchers have developed the first end-to-end text-to-speech (TTS) system for Efik, a low-resource tonal language spoken in Nigeria. The study involved creating a corpus of 2,632 utterances and evaluating four neural models: VITS, MMS-TTS, SpeechT5, and Orpheus-TTS. MMS-TTS performed best, achieving a MOS of 3.80, though it still exhibited tonal errors. The findings underscore the need for larger datasets and tone-aware modeling for African languages. AI
IMPACT This research provides a baseline for developing speech synthesis technologies for underrepresented African languages, potentially improving accessibility and digital preservation.
RANK_REASON Academic paper detailing a new TTS system for a low-resource language.
- arXiv
- Efik
- MMS-TTS
- Nat-MOS
- Nigeria
- Orpheus-TTS
- SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing
- VITS
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →