---
license: apache-2.0
language: [wuu]
tags: [automatic-speech-recognition, onnx, onnx-asr, whisper, wu-chinese]
base_model: kaiwang0574/whisper-wu
---

# whisper-wu — ONNX

ONNX export of [kaiwang0574/whisper-wu](https://huggingface.co/kaiwang0574/whisper-wu)
(a LoRA adapter over `openai/whisper-small`, fine-tuned for Wu Chinese by
[kaiwang0574](https://huggingface.co/kaiwang0574)) for
[onnx-asr](https://github.com/istupakov/onnx-asr) (standard `whisper` model type — works with
stock onnx-asr, no patches needed). fp32 and int8 variants included.

The source repo only publishes a PEFT/LoRA adapter (`adapter_config.json` +
`adapter_model.safetensors`, `base_model_name_or_path: openai/whisper-small`), not a merged
checkpoint. This export merges the adapter into the base model (`peft.merge_and_unload()`)
before running the standard ONNX export pipeline.

License: apache-2.0, inherited from the source model.

First specialized ONNX ASR model for Wu Chinese in this collection.

## Usage

Whisper has no dedicated `wuu` language token; use `language="zh"` — the Wu-dialect behavior
comes from the fine-tune itself.

```python
import onnx_asr
model = onnx_asr.load_model("whisper", "path/to/this/repo")  # or quantization="int8"
print(model.recognize("audio_16khz.wav", language="zh"))
```

Verified on a clip from [MagicHub/magicdata-dialect-wu-chinese-tts-lite](https://huggingface.co/datasets/MagicHub/magicdata-dialect-wu-chinese-tts-lite)
(Jiangsu dialect corpus; FLEURS does not cover Wu Chinese):

- Reference: 倷今朝阿去园林里白相？里向个荷花开的交关好看。
- fp32 (RTF 0.21): 內徑在安赤與林立巴線裡,只剩個好伙計的接歸喊口。
- int8 (RTF 0.17): 內境在安赤與林立巴線,只像個好火開的接歸駭口。

**Verified with caveats, honestly**: the ONNX fp32 output was cross-checked against the
merged model run natively through `transformers` on the same clip, and the two match closely
("內徑在安赤與林立巴線裡,只剩下個好火開的接歸客戶。") — confirming the ONNX conversion is
faithful. The transcript itself, however, diverges substantially from the reference text.
This is a real accuracy limitation of the small LoRA adapter (rank 16, `q/k/v/o_proj` only)
on this low-resource dialect, not an artifact of this export. Treat this as a correctly
converted mirror of the source model, not a claim of strong Wu Chinese accuracy. RTF measured
on an AMD Ryzen 5 7600 (CPU, 4 OMP threads, shared/loaded box — not a clean benchmark number).

Int8 decoder was produced by quantizing the pre-merge decoders separately and re-merging
(`merge_decoders(..., strict=False)`); direct quantization of the merged decoder graph does
not shrink it (its `If` subgraphs are skipped by onnxruntime's dynamic quantizer).