Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- zh
|
| 5 |
+
base_model: inclusionAI/Ming-3.0-tts-0.5B
|
| 6 |
+
tags:
|
| 7 |
+
- text-to-speech
|
| 8 |
+
- on-device
|
| 9 |
+
- onnx
|
| 10 |
+
- int4
|
| 11 |
+
- distillation
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# WebFM Ming-3.0-tts 0.5B
|
| 15 |
+
|
| 16 |
+
Experimental WebFM assets for replacing F5-TTS with
|
| 17 |
+
[`inclusionAI/Ming-3.0-tts-0.5B`](https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B).
|
| 18 |
+
|
| 19 |
+
This repository currently contains:
|
| 20 |
+
|
| 21 |
+
- `distill/dit4.onnx`: 4-step CFG-free flow DiT student (fp32 ONNX).
|
| 22 |
+
- `distill/dit4_int4.onnx`: the same student quantized with ONNX Runtime
|
| 23 |
+
`MatMulNBits` (4-bit weights, block size 32, symmetric, accuracy level 4).
|
| 24 |
+
- `distill/student.pt`: PyTorch state dict for the student DiT.
|
| 25 |
+
- `samples/`: before/after listening examples.
|
| 26 |
+
- `samples_report.json`: durations and int4 validation metrics.
|
| 27 |
+
|
| 28 |
+
## Listening examples
|
| 29 |
+
|
| 30 |
+
Text: `雨落在窗台上，远处的灯一盏一盏亮起来。`
|
| 31 |
+
|
| 32 |
+
| Stage | File | Duration |
|
| 33 |
+
| --- | --- | ---: |
|
| 34 |
+
| Teacher, 10-step CFG=2 | [teacher_10step.wav](samples/teacher_10step.wav) | 3.70 s |
|
| 35 |
+
| Student, 4-step CFG-free | [student_4step.wav](samples/student_4step.wav) | 3.68 s |
|
| 36 |
+
| Student + int4 DiT | [student_4step_dit_int4.wav](samples/student_4step_dit_int4.wav) | 3.68 s |
|
| 37 |
+
|
| 38 |
+
The int4 sample uses the PyTorch LM, aggregator and vocoder with the quantized
|
| 39 |
+
ONNX DiT. Full LM/vocoder/encoder quantization is in progress.
|
| 40 |
+
|
| 41 |
+
## Distillation
|
| 42 |
+
|
| 43 |
+
- Teacher: 10 Euler steps, CFG=2, fp32 on Tesla T4.
|
| 44 |
+
- Student: 4 Euler steps, CFG-free, initialized from the teacher DiT.
|
| 45 |
+
- Data: 200 SFW Chinese utterances, 13,112 latent frames, generated from
|
| 46 |
+
`books/story.txt` with the `lingguang` prompt voice.
|
| 47 |
+
- Training: 2,000 optimizer steps, batch 16, AdamW lr 1e-4, fp32.
|
| 48 |
+
- Loss: `1.0234 -> 0.2379`.
|
| 49 |
+
- Student ONNX parity: `max_abs=1.19e-6` against PyTorch.
|
| 50 |
+
- Int4 DiT: cosine `0.9962`, `max_abs=0.5156` on a random probe.
|
| 51 |
+
|
| 52 |
+
## Status
|
| 53 |
+
|
| 54 |
+
This is experimental research output, not a WebFM release model. The full
|
| 55 |
+
browser bundle (LM, aggregator, encoder, vocoder, host tensors) and the WebGPU
|
| 56 |
+
runtime are not included yet.
|
| 57 |
+
|
| 58 |
+
## License and safety
|
| 59 |
+
|
| 60 |
+
The base model is MIT-licensed. The original usage disclaimer applies: label
|
| 61 |
+
synthetic speech as synthetic, do not impersonate a real person without consent,
|
| 62 |
+
and do not use the model for deception or illegal purposes. This release is SFW
|
| 63 |
+
only.
|