Model Card for ZipVoice-Distill (IB-Robot)

ZipVoice-Distill text-to-speech (flow-matching decoder + Vocos vocoder) packaged for the IB-Robot framework with two deployments sharing one contract.

Deployments

deployment	backend	artifacts	notes
ascend_310p	Ascend ACL (Ascend310P1)	text_encoder + flow_decoder OM	fixed buckets (t256 / f1537), CFG-distilled 4-step
ubuntu_onnx	ONNX Runtime CPU (torch profile)	dynamic-shape ONNX pair	host-side synthesis, 8-step default

Shared contract: tensor_model/zipvoice/synthesize (tts.text/prompt_audio/prompt_sample_rate/prompt_text -> tts.audio), preprocessing emilia-zh-cn2an-jieba-pypinyin-fixed-golden-prompt-v1, output mono-float32-pcm-24000hz-vocos-cpu.

Provenance note

• ONNX weights re-exported from the official k2-fsa/ZipVoice ModelScope release (zipvoice_distill), sha256 pinned in the manifest artifacts.
• The 310P OM pair was converted from the same model with fixed shapes (text_capacity=256, flow_frames=1537, original logical revision zipvoice-distill-310p1-bucket-2026-08-03).
• Shared assets (tokens.txt, vocos/pytorch_model.bin, prompts/default.npz) are byte-identical across both deployments.
• Inference hyperparameters differ by design: 310P uses 4 diffusion steps, the host ONNX path defaults to 8 (see assets/zipvoice_310p.json vs assets/zipvoice_onnx.json).

Repository Structure

• inference_manifest.json — deployment routing (schema v3, iterative request contract)
• assets/zipvoice_310p.json / assets/zipvoice_onnx.json — per-deployment runtime config
• assets/tokens.txt, assets/vocos/, assets/prompts/ — shared assets
• artifacts/ascend/ascend_310p/*.om — compiled fixed-shape modules
• artifacts/onnx/*.onnx — dynamic-shape ONNX exports

Usage

Host: voice_tts_service with bundle_path: models/zipvoice, deployment: ubuntu_onnx. Board: select ascend_310p through the unified inference runtime.

License

Code and packaging: Apache-2.0. ZipVoice model weights follow the k2-fsa/ZipVoice upstream license.