README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-ASR-0.6B
pipeline_tag: automatic-speech-recognition
language:
- zh
- en
- fr
- de
- ja
- ko
- yue
tags:
- speech-recognition
- distillation
- multilingual
---

# asr-324m-apache

A 7-language (Chinese, English, French, German, Japanese, Korean, Cantonese) speech-to-text model
distilled from [Qwen3-ASR-0.6B](https://huggingface.co/Qwen/Qwen3-ASR-0.6B) (Apache-2.0), depth- and
vocabulary-pruned to **323.77M parameters** — matching
[Audio8-ASR-0.1B](https://huggingface.co/Audio8/Audio8-ASR-0.1B)'s size, an existing model covering
the same languages that is CC-BY-NC and cannot be used commercially. This checkpoint is fully
Apache-2.0.

Base checkpoint (pre-compression, 467.81M, ties Audio8's quality): [`Luigi/asr-468m-apache-base`](https://huggingface.co/Luigi/asr-468m-apache-base).
Code, full training pipeline, and every finding (including two negative results on adding more
training data): [github.com/vieenrose/asr-324m-apache](https://github.com/vieenrose/asr-324m-apache).

## Results (200-clip FLEURS test gate, all-refs; CER for zh/ja/ko/yue, WER for en/fr/de)

| language | Audio8-ASR-0.1B | this model | delta |
|---|---|---|---|
| French | 20.88 | **19.29** | −1.59 ✅ |
| Korean | 13.97 | **12.47** | −1.50 ✅ |
| Japanese | 17.97 | **16.64** | −1.33 ✅ |
| German | 17.73 | 19.46 | +1.73 |
| English | 8.51 | 12.36 | +3.85 |
| Chinese | 11.94 | 18.00 | +6.06 |
| Cantonese | 16.16 | 22.54 | +6.38 |
| **macro** | **15.31** | **17.25** | **+1.94** |

3 of 7 languages individually beat Audio8 at this size; the overall average does not, yet — the
remaining gap is a data-domain problem in Chinese and Cantonese specifically, not an architecture
limit (see the linked repo's `docs/findings.md`).

## Usage

This checkpoint's vocabulary is pruned (151,936 → 42,000 ids), so decoding needs a remap step —
`vocab_remap.json` (shipped in this repo) maps between the model's compact id space and the
original Qwen3 tokenizer's ids.

```python
import json
import torch
from huggingface_hub import hf_hub_download
from qwen_asr.core.transformers_backend.modeling_qwen3_asr import Qwen3ASRForConditionalGeneration
from qwen_asr.core.transformers_backend.processing_qwen3_asr import Qwen3ASRProcessor

path = "Luigi/asr-324m-apache"
proc = Qwen3ASRProcessor.from_pretrained(path)
model = Qwen3ASRForConditionalGeneration.from_pretrained(path, dtype=torch.bfloat16).cuda().eval()

remap = json.load(open(hf_hub_download(path, "vocab_remap.json")))
old_to_new = {int(k): v for k, v in remap["old_to_new"].items()}
keep_ids = remap["keep_ids"]
IM_END = 151645  # Qwen3 tokenizer's original <|im_end|> id

def transcribe(wav_16k_float32, language="Chinese", max_new_tokens=128):
    NATIVE = ("<|im_start|>system\n<|im_end|>\n<|im_start|>user\n<|audio_pad|><|im_end|>\n"
              "<|im_start|>assistant\n")
    e = proc(text=NATIVE + f"language {language}<asr_text>", audio=[wav_16k_float32],
             sampling_rate=16000, return_tensors="pt")
    e["input_ids"] = e["input_ids"].apply_(lambda i: old_to_new[i])
    eos = pad = old_to_new[IM_END]
    e = {k: (v.cuda() if torch.is_tensor(v) else v) for k, v in e.items()}
    if "input_features" in e:
        e["input_features"] = e["input_features"].to(torch.bfloat16)
    with torch.no_grad(), torch.autocast("cuda", dtype=torch.bfloat16):
        out = model.generate(**e, max_new_tokens=max_new_tokens, do_sample=False,
                             eos_token_id=eos, pad_token_id=pad)
    ids = out[0][e["input_ids"].shape[1]:].tolist()
    ids = [keep_ids[i] for i in ids]                    # map back to the original vocab space
    return proc.tokenizer.decode(ids, skip_special_tokens=True)
```

`language` accepts: Chinese, English, French, German, Japanese, Korean, Cantonese.

## Training data and attribution

Trained on Common Voice 17 (CC0), WenetSpeech4TTS, Multilingual LibriSpeech, LibriSpeech, and FLEURS
(all CC-BY-4.0). This model was trained in part on WenetSpeech4TTS, Multilingual LibriSpeech,
LibriSpeech, and FLEURS, each licensed CC-BY-4.0 by their respective creators. Full source breakdown
and every negative data-augmentation result are in the linked GitHub repository.

Audio8-ASR-0.1B is used only as a measurement reference throughout — its weights are never loaded,
merged, or distilled from.