Source: https://mayphus.org/wuxue-speech/
Title: Wuxue speech recognition
Metadata: {"category":"research","id":"wuxue-speech","kind":"research","language":"en","route":"/wuxue-speech/","tags":["speech recognition","Wuxue","dialect","machine learning"],"type":"project"}

# Wuxue speech recognition

A living project record of speech recognition for Wuxue dialect. The current experiment tests a small adaptation workflow; it does not establish broad Wuxue recognition capability.

## Current experiment

The reviewed collection contains **44 clips totaling 139.16 seconds**. **36 clips are used for training; eight are excluded from training and used for exploratory evaluation.** The same labels and split are used across the comparisons below.

The starting model is multilingual `openai/whisper-tiny`. Training completed 200 steps. Checkpoint saving, resuming, and reloading were verified.

## Aggregate results

Character error rate (CER) measures character substitutions, deletions, and insertions relative to the reference character count. **Lower is better.** CER can exceed 100% when insertions increase the error count beyond the reference length. These results use the same text normalization.

| Model state | Training CER | Exploratory evaluation CER |
| --- | ---: | ---: |
| Pretrained base | 87.01% | 126.67% |
| Step 40 | 51.95% | 85.00% |
| Step 200 | 2.60% | 70.00% |

The training error falls much further than the evaluation error. This gap is consistent with substantial overfitting on a tiny collection. It does not establish reliable recognition across speakers or Wuxue speech generally.

The eight evaluation clips have been reused during exploration. They are **not an independent final benchmark**, and the results should not be read as a generalization claim.

## Frozen-model comparison

Three pretrained models have now been evaluated without further fine-tuning on the **same eight exploratory evaluation clips**. The adapted Whisper tiny result is included as a reference, not as a frozen model.

| Model | Adaptation in this project | Exploratory evaluation CER |
| --- | --- | ---: |
| Fun-ASR-Nano-2512 | None; frozen pretrained model | 58.33% |
| Qwen3-ASR-0.6B | None; frozen pretrained model | 68.33% |
| Multilingual Whisper small | None; frozen pretrained model | 105.00% |
| Multilingual Whisper tiny, step 200 | Earlier 200-step fine-tuning | 70.00% |

Fun-ASR-Nano had the lowest CER in this comparison. These are results on a tiny, reused evaluation set, **not evidence of broad Wuxue recognition accuracy or a reliable ranking beyond these clips**. No additional fine-tuning was performed.

The comparison required a correction to the Nano model-loading path and used approximately **0.12 GPU-hours** in total.

### Tested models and public references

- **[Fun-ASR-Nano-2512](https://huggingface.co/FunAudioLLM/Fun-ASR-Nano-2512)** — 800M parameters, Apache 2.0. Its model card advertises Gan dialect and Hubei-accent coverage, which is not a Wuxue-specific guarantee. The official FunASR repository provides [LoRA fine-tuning code](https://github.com/modelscope/FunASR/blob/main/examples/industrial_data_pretraining/fun_asr_nano/lora_finetune.sh), despite the model card's training TODO.
- **[Qwen3-ASR-0.6B](https://huggingface.co/Qwen/Qwen3-ASR-0.6B)** — Apache 2.0. Its advertised Chinese dialect/accent categories include Hubei and Jiangxi. [Official full-model fine-tuning code](https://github.com/QwenLM/Qwen3-ASR/tree/main/finetuning) is available; feasibility of fine-tuning on a T4 has not been verified here.
- **[Multilingual Whisper small](https://github.com/openai/whisper)** — 244M parameters, a larger same-family control for the tiny baseline.

## Next steps

Use this exploratory comparison to inform model selection before deciding on further adaptation. A separate, independent evaluation is still needed before making broader performance claims. Future updates will distinguish verified milestones from plans and retain the evaluation caveats.

## Public record and private data

This page shares methods, model names, aggregate counts, and aggregate results only. Voice recordings, transcripts, example utterances, predictions, speaker identifiers, private working links, and model checkpoints are not published here.