---
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-468m-apache-base

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). **This is
the Stage-1 base checkpoint** — the quality target reached before any parameter compression, at
467.81M parameters. It statistically ties
[Audio8-ASR-0.1B](https://huggingface.co/Audio8/Audio8-ASR-0.1B) (macro 15.36 vs 15.31) at ~1.4x its
size, fully Apache-2.0 where Audio8 is CC-BY-NC and unusable commercially.

**If you want the smaller, size-matched release** (323.77M, Audio8's exact parameter budget, at a
quality cost — see its model card for the honest tradeoff), use
[`Luigi/asr-324m-apache`](https://huggingface.co/Luigi/asr-324m-apache) instead. This base
checkpoint is also the required starting point for reproducing that model's compression pipeline.

Code, full training pipeline, and every finding: [github.com/vieenrose/asr-324m-apache](https://github.com/vieenrose/asr-324m-apache).

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

**15.36** vs Audio8-ASR-0.1B's 15.31 — a statistical tie.

## Usage

Unlike the 324M release, this checkpoint's vocabulary is **not** pruned (full 151,936-id Qwen3
tokenizer), so no id remapping is needed.

```python
import torch
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-468m-apache-base"
proc = Qwen3ASRProcessor.from_pretrained(path)
model = Qwen3ASRForConditionalGeneration.from_pretrained(path, dtype=torch.bfloat16).cuda().eval()

NATIVE = ("<|im_start|>system\n<|im_end|>\n<|im_start|>user\n<|audio_pad|><|im_end|>\n"
          "<|im_start|>assistant\n")

def transcribe(wav_16k_float32, language="Chinese", max_new_tokens=128):
    e = proc(text=NATIVE + f"language {language}<asr_text>", audio=[wav_16k_float32],
             sampling_rate=16000, return_tensors="pt")
    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)
    ids = out[0][e["input_ids"].shape[1]:].tolist()
    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.

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