Update model card: multi-deployment (torch/onnx) support
Browse files
README.md
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
pipeline_tag: text-to-speech
|
| 4 |
+
tags:
|
| 5 |
+
- zipvoice
|
| 6 |
+
- tts
|
| 7 |
+
- speech-synthesis
|
| 8 |
+
- robotics
|
| 9 |
+
- IB-Robot
|
| 10 |
+
- ascend
|
| 11 |
+
- onnx
|
| 12 |
+
- edge-deployment
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Model Card for ZipVoice-Distill (IB-Robot)
|
| 16 |
+
|
| 17 |
+
ZipVoice-Distill text-to-speech (flow-matching decoder + Vocos vocoder)
|
| 18 |
+
packaged for the [IB-Robot](https://atomgit.com/openeuler/IB_Robot) framework
|
| 19 |
+
with two deployments sharing one contract.
|
| 20 |
+
|
| 21 |
+
## Deployments
|
| 22 |
+
|
| 23 |
+
| deployment | backend | artifacts | notes |
|
| 24 |
+
|---|---|---|---|
|
| 25 |
+
| `ascend_310p` | Ascend ACL (Ascend310P1) | text_encoder + flow_decoder OM | fixed buckets (t256 / f1537), CFG-distilled 4-step |
|
| 26 |
+
| `ubuntu_onnx` | ONNX Runtime CPU (torch profile) | dynamic-shape ONNX pair | host-side synthesis, 8-step default |
|
| 27 |
+
|
| 28 |
+
Shared contract: `tensor_model/zipvoice/synthesize`
|
| 29 |
+
(`tts.text/prompt_audio/prompt_sample_rate/prompt_text -> tts.audio`),
|
| 30 |
+
preprocessing `emilia-zh-cn2an-jieba-pypinyin-fixed-golden-prompt-v1`, output
|
| 31 |
+
`mono-float32-pcm-24000hz-vocos-cpu`.
|
| 32 |
+
|
| 33 |
+
## Provenance note
|
| 34 |
+
|
| 35 |
+
- ONNX weights re-exported from the official k2-fsa/ZipVoice ModelScope
|
| 36 |
+
release (`zipvoice_distill`), sha256 pinned in the manifest artifacts.
|
| 37 |
+
- The 310P OM pair was converted from the same model with fixed shapes
|
| 38 |
+
(`text_capacity=256`, `flow_frames=1537`, original logical revision
|
| 39 |
+
`zipvoice-distill-310p1-bucket-2026-08-03`).
|
| 40 |
+
- Shared assets (`tokens.txt`, `vocos/pytorch_model.bin`,
|
| 41 |
+
`prompts/default.npz`) are byte-identical across both deployments.
|
| 42 |
+
- Inference hyperparameters differ by design: 310P uses 4 diffusion steps,
|
| 43 |
+
the host ONNX path defaults to 8 (see `assets/zipvoice_310p.json` vs
|
| 44 |
+
`assets/zipvoice_onnx.json`).
|
| 45 |
+
|
| 46 |
+
## Repository Structure
|
| 47 |
+
|
| 48 |
+
- `inference_manifest.json` — deployment routing (schema v3, iterative request contract)
|
| 49 |
+
- `assets/zipvoice_310p.json` / `assets/zipvoice_onnx.json` — per-deployment runtime config
|
| 50 |
+
- `assets/tokens.txt`, `assets/vocos/`, `assets/prompts/` — shared assets
|
| 51 |
+
- `artifacts/ascend/ascend_310p/*.om` — compiled fixed-shape modules
|
| 52 |
+
- `artifacts/onnx/*.onnx` — dynamic-shape ONNX exports
|
| 53 |
+
|
| 54 |
+
## Usage
|
| 55 |
+
|
| 56 |
+
Host: `voice_tts_service` with `bundle_path: models/zipvoice`,
|
| 57 |
+
`deployment: ubuntu_onnx`. Board: select `ascend_310p` through the unified
|
| 58 |
+
inference runtime.
|
| 59 |
+
|
| 60 |
+
## License
|
| 61 |
+
|
| 62 |
+
Code and packaging: Apache-2.0. ZipVoice model weights follow the k2-fsa/ZipVoice
|
| 63 |
+
upstream license.
|