MMS-TTS Dhivehi — Quantized checkpoints
Locally quantized (dynamic int8) PyTorch state-dicts of the Dhivehi MMS-TTS (VITS)
checkpoints, produced during the Dhivehi TTS project's quantization campaign
(quantize_mms.py, chatterbox-tts-dhivehi).
Base weights (not ours): the six models below were published by alakxender (MIT license). This repo contains our derived quantization artifacts and is republished with attribution.
| File | Base checkpoint (alakxender) |
|---|---|
mms-tts-div-finetuned-md-f01.pt |
alakxender/mms-tts-div-finetuned-md-f01 |
mms-tts-div-finetuned-md-f02.pt |
alakxender/mms-tts-div-finetuned-md-f02 |
mms-tts-div-finetuned-md-f03.pt |
alakxender/mms-tts-div-finetuned-md-f03 |
mms-tts-div-finetuned-md-m01.pt |
alakxender/mms-tts-div-finetuned-md-m01 |
mms-tts-div-ft-spk01-f01.pt |
alakxender/mms-tts-div-ft-spk01-f01 |
mms-tts-div-ft-spk01-m01.pt |
alakxender/mms-tts-div-ft-spk01-m01 |
Honest caveats
- Dynamic INT8 quantizes only the
nn.Linearlayers; MMS/VITS has just four of them, so these files are ~142.8 MB vs the ~145 MB FP32 originals — functional, but not a major compression or RAM win. - All six reload and generate valid Dhivehi audio (validated on the GTX 1650 benchmark rig). No listening panel was run; perceptual quality is assumed near-identical because most of the model was left at FP32.
Related
- ONNX exports (fp32 + int8):
mohamedrayyan/mms-tts-dhivehi-onnx - Dhivehi TTS project incl. benchmark reports:
mohamedrayyan/chatterbox-tts-dhivehi