Qwen3-ASR-0.6B — Korean 119 emergency-call LoRA (tiny)
A LoRA adapter for the smaller Qwen/Qwen3-ASR-0.6B-hf, same domain and training recipe as Playedwell03/qwen3-asr-1.7b-119ko-v5, made for CPU/edge deployment where the 1.7B model is too heavy (e.g. ONNX export to a low-power device).
Results vs the 1.7B adapter
Same held-out 500-utterance test, normalized CER (corpus-level):
| Condition | 0.6B base | this (0.6B tuned) | 1.7B tuned (v5) |
|---|---|---|---|
| Clean | 0.331 | 0.187 | 0.188 |
| Siren, 5dB | — | 0.231 | 0.184 |
| Background noise, 10dB | — | 0.262 | 0.217 |
| Crowd babble, 10dB | — | 0.319 | 0.245 |
| Background noise, 0dB | — | 0.458 | 0.344 |
| Crowd babble, 5dB | — | 0.459 | 0.336 |
| Crowd babble, 0dB | — | 0.673 | 0.597 |
On clean audio this 0.6B adapter is statistically indistinguishable from the 1.7B one (diff -0.001, 95% CI [-0.030, +0.029]) despite having a third of the parameters — the domain-adaptation gain from fine-tuning was in fact larger for this smaller model (-0.144 vs -0.110 for the 1.7B).
Under noise, this trade-off reverses: the 1.7B adapter is significantly better in 6 of 7 tested noise conditions (95% CI excludes 0 for all but 10dB background noise, which is borderline). The gap is largest at moderate noise (crowd babble 5dB: -0.124; background noise 0dB: -0.114). Hallucination suppression on noise-only input (0% output, n=50 per condition) holds equally well for both sizes.
Use the 1.7B adapter if noise robustness matters for your deployment. Use this one only where the environment is expected to be quiet, or where the compute budget genuinely doesn't allow 1.7B.
Training
Identical recipe to the 1.7B adapter except for LoRA sizing, tuned against two independent public Qwen3-ASR-0.6B LoRA examples (AleksanderObuchowski/Qwen3-ASR-0.6B-med-pl-lora-decoder-only, SayedShaun/Qwen3-ASR-0.6B-Bengali-LoRA) rather than just copying the 1.7B settings down:
- LoRA: r=32, alpha=32 (1:1, matching both reference adapters — the 1.7B adapter used a 2:1 ratio with no particular justification), dropout=0.05
- target_modules: decoder attention+MLP (
q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj) plusout_proj,fc1,fc2for the audio encoder, which uses different names for its output projection and MLP than the decoder does — without these three the encoder is only partially adapted (verified against the model's source:Qwen3ASRAudioAttention/Qwen3ASRAudioEncoderLayer) - Same data (20h, AI Hub 119 calls), same noise augmentation (47%/6%/47% split, 0–15dB SNR), same optimizer/schedule/epochs as the 1.7B adapter — see that model card for details
- 29.5M trainable params / 811.9M total (3.6%)
Known limitations
Same caveats as the 1.7B adapter (in-domain-only test, synthetic noise, hotword prompting hurts performance, no human-in-the-loop substitute) — see Playedwell03/qwen3-asr-1.7b-119ko-v5 for the full list. Additionally: not validated after ONNX export or INT8 quantization — the numbers above are PyTorch/bf16 only.
Usage
import torch
from peft import PeftModel
from transformers import AutoProcessor, Qwen3ASRForConditionalGeneration
BASE = "Qwen/Qwen3-ASR-0.6B-hf"
ADAPTER = "Playedwell03/qwen3-asr-0.6b-119ko-tiny"
processor = AutoProcessor.from_pretrained(BASE)
model = Qwen3ASRForConditionalGeneration.from_pretrained(BASE, dtype=torch.bfloat16, device_map="cuda")
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()
inputs = processor.apply_transcription_request(audio="call.wav", language="Korean").to(model.device, model.dtype)
output_ids = model.generate(**inputs, max_new_tokens=256)
generated = output_ids[:, inputs["input_ids"].shape[1]:]
print(processor.decode(generated, return_format="transcription_only")[0])
Requires transformers>=5.13.0 and peft.