SimpleKhmerTTS
Decoder-only transformer TTS, using WavTokenizer as the audio codec and DDD-Cambodia/khmer-speech-dataset as dataset.
Samples
| Text | Audio |
|---|---|
| Sample 1 | |
| Sample 2 | |
| Sample 3 | |
| Sample 4 |
Setup
git clone https://github.com/FirstPotatoCoder/SimpleKhmerTTS.git
cd SimpleKhmerTTS
pip install -q -r requirements.txt
python scripts/download_weights.py
Run
python examples/run_inference.py " ភាសាខ្មែរមានអក្សរច្រើន ហើយពិបាកសរសេរបន្តិច។ ភាសានេះមានប្រវត្តិយូរអង្វែង និងជាផ្នែកមួយយ៉ាងសំខាន់នៃវប្បធម៌កម្ពុជា។" --speaker_id 0
Or from Python:
from tts.inference import TTSPipeline
pipe = TTSPipeline(
tts_weights="weights/tts.pt",
wavtokenizer_weights="weights/wavtokenizer.ckpt",
wavtokenizer_config="configs/wavtokenizer_config.yaml"
)
pipe.generate(" ភាសាខ្មែរមានអក្សរច្រើន ហើយពិបាកសរសេរបន្តិច។", speaker_id=1, out_path="out.wav")
Limitations
- The model sometimes hallucinates, generating babbling or garbled output on rare or unseen words, making it ill-suited for direct deployment — likely due to the limited amount of training data and codec's limited performance on Khmer audio.
- Works best when generating ~3s to 10s of audio, matching the length distribution of its training data.
- No chunking support yet — the current repo only supports clip-by-clip generation, one sample at a time.
Credits
-
Data: Khmer Speech Dataset by DDD-Cambodia
-
Compute: Kaggle for free T4 GPU access
-
Audio tokenizer/codec: WavTokenizer (vendored under
wavtokenizer/, trimmed to inference-only code).