MMS-TTS Dhivehi — ONNX exports
ONNX Runtime exports (FP32 + dynamic INT8) of the Dhivehi MMS-TTS (VITS) checkpoints, produced locally in the Dhivehi TTS project for dependency-free CPU deployment.
Base weights (not ours): exported from the models published by alakxender (MIT). This repo contains our derived export/quantization work, republished with attribution.
| File | Base checkpoint (alakxender) |
|---|---|
mms-tts-div-finetuned-md-f01.onnx / .int8.onnx |
alakxender/mms-tts-div-finetuned-md-f01 |
mms-tts-div-finetuned-md-f02.onnx / .int8.onnx |
alakxender/mms-tts-div-finetuned-md-f02 |
mms-tts-div-finetuned-md-f03.onnx / .int8.onnx |
alakxender/mms-tts-div-finetuned-md-f03 |
mms-tts-div-finetuned-md-m01.onnx / .int8.onnx |
alakxender/mms-tts-div-finetuned-md-m01 |
mms-tts-div-ft-spk01-f01.onnx / .int8.onnx |
alakxender/mms-tts-div-ft-spk01-f01 |
mms-tts-div-ft-spk01-m01.onnx / .int8.onnx |
alakxender/mms-tts-div-ft-spk01-m01 |
Measured numbers (GTX 1650 rig, one warm sample each)
| Mode | Size | Real-time factor (CPU) |
|---|---|---|
| ONNX FP32 | ~114 MB | 0.157 — faster than the PyTorch CPU baseline |
| ONNX dynamic INT8 | ~38 MB | 1.77 — smaller but slower than real time in this graph |
FP32 is the deployment pick for CPU speed; INT8 wins on disk/RAM footprint. All 12 graphs passed finite-audio and two-length sanity checks. The export used legacy tracing, so treat the graphs as fixed-architecture artifacts, not a general exporter.
Related
- Quantized PyTorch checkpoints:
mohamedrayyan/mms-tts-dhivehi-quantized - Dhivehi TTS project incl. benchmark reports:
mohamedrayyan/chatterbox-tts-dhivehi