ProjectSpeech AI / October 2026In progress
Khmer TTS
- Question
- Can a compact English text-to-speech model learn to read Khmer reliably, clone a voice from a short prompt, and still run on a laptop CPU?
- Approach
- Fine-tune Kyutai's Pocket TTS on licensed Khmer speech with a new tokenizer and a text front-end for numbers and mixed scripts, judged by automatic metrics and a native listener.
- Status
- Teaching a compact English text-to-speech model to read Khmer, on a small budget, with a native speaker as the final judge.
- Built with
- Python · PyTorch · Pocket TTS · Mimi
What I am building
- Data: licence-checked Khmer read speech, cleaned conservatively, with forced alignment to cut voice prompts at natural pauses.
- Text front-end: a number reader for dates, times, money and percentages, with every rule checked against real recordings, and handling for mixed Khmer and Latin text.
- Training: fine-tuning Kyutai’s Pocket TTS with a new Khmer tokenizer, each experiment planned in advance and compared on the same held-out speakers.
- Evaluation: automatic intelligibility and voice-similarity scores, plus blinded listening by a native speaker, because automatic judges miss pronunciation errors a listener hears. The evaluation post explains why.
Status
A working system exists. The data preparation and training are written up in Same model, new language: fine-tuning Pocket TTS to speak Khmer. Inference, the text front-end and pronunciation fixes will follow.