LoRA adapter for Qwen/Qwen3-ASR-1.7B fine-tuned on Synthetic K-OR Speech Audio v1, a Korean operating-room speech corpus. Targets the OR-specific phonetic Korean transliterations of English clinical vocabulary (보비 / 석션 / 로큐 / 슈감마 / 메첸바움 / 이티씨오투, etc.) that off-the-shelf ASR systems struggle with.

Results

Test split: 800 clips (200 utterances × 4 voices, utterance-id stratified, seed 20260503).

ASR	CER	vs Whisper baseline
Qwen3-ASR-1.7B + this adapter	0.0334	-59.8%
Whisper-v3-turbo + LoRA	0.0431	-48.0%
Whisper-v3-turbo (baseline)	0.0829	—
Qwen3-ASR-1.7B (baseline)	0.0870	+4.9%

CER by code-switching style (where the gap matters most):

cs_style	Baseline	This adapter
`none` (pure Korean)	0.046	0.018
`phonetic_kr` (음차)	0.078	0.022
`mixed` (KR+EN)	0.140	0.067
`english`	0.473	0.276

Usage

pip install qwen-asr peft

python

import torch
from peft import PeftModel
from qwen_asr.core.transformers_backend.modeling_qwen3_asr import Qwen3ASRForConditionalGeneration
from transformers import AutoProcessor

base = "Qwen/Qwen3-ASR-1.7B"
adapter = "vitaldb/qwen3-asr-1.7b-kor-or-lora"

processor = AutoProcessor.from_pretrained(base, trust_remote_code=True)
model = Qwen3ASRForConditionalGeneration.from_pretrained(base, dtype=torch.bfloat16)
model.thinker = PeftModel.from_pretrained(model.thinker, adapter)
model.to("cuda").eval()

import librosa
y, _ = librosa.load("clip.wav", sr=16000, mono=True)

msgs = [{"role":"system","content":""}, {"role":"user","content":[{"type":"audio","audio":""}]}]
prompt = processor.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
prompt += "language Korean<asr_text>"

inputs = processor(text=[prompt], audio=[y], return_tensors="pt", padding=True)
inputs = {k: v.to("cuda") for k, v in inputs.items()}
inputs["input_features"] = inputs["input_features"].to(torch.bfloat16)

with torch.no_grad():
    gen = model.generate(**inputs, max_new_tokens=128)
new_tokens = gen.sequences[:, inputs["input_ids"].shape[1]:]
text = processor.tokenizer.batch_decode(new_tokens, skip_special_tokens=True)[0].strip()
print(text)

Training

Base model: Qwen/Qwen3-ASR-1.7B (Apache-2.0)
Trainable params: 8.78M / 2.05B (0.43%)
Target modules: q_proj, k_proj, v_proj, o_proj (text decoder; audio encoder frozen)
LoRA rank / alpha / dropout: 16 / 32 / 0.05
Optimizer: AdamW, learning rate 1e-4, warmup 100 steps
Schedule: 3 epochs, batch 2 × grad_accum 8 (effective 16), bf16, gradient checkpointing
Hardware: NVIDIA RTX 4090 24 GB
Train time: ~82 minutes
Train data: 6,400 clips (1,600 utterances × 4 voices) from Synthetic K-OR Speech Audio v1

Dataset

The adapter is trained on Synthetic K-OR Speech Audio v1 — 8,000 audio clips (2,000 utterances × 4 voice profiles) synthesized from Synthetic K-OR Speech Corpus v1.0 text via Qwen3-TTS.

Voice profiles:

chul — cloned voice of an SNUH anesthesiologist (male, written consent)
sohee — preset Korean female (warm)
ethan — preset multilingual male, intent-conditioned prosody
maia — preset multilingual female, intent-conditioned prosody

Companion text corpus: Synthetic K-OR Speech Corpus v1.0 (CC-BY-4.0).

Limitations

Trained on synthetic TTS audio, not real OR recordings. Real-environment transfer (overlap, machine noise, reverberation) is not validated.
Single-institution lexicon (SNUH conventions). Other Korean tertiary hospitals may use different transliteration variants.
Single male reference voice for cloning; gender/dialect diversity comes from the three preset voices.
The adapter inherits Qwen3-TTS acoustic characteristics. Apply with caution to truly out-of-distribution audio.

License

Apache-2.0 (matching base model).

The adapter weights themselves are unrestricted. The training dataset (Synthetic K-OR Speech Audio v1) is also Apache-2.0 with a use restriction prohibiting reuse for TTS / voice-cloning / speaker-ID development.

Citation

bibtex

@misc{kor_or_qwen3_asr_lora_2026,
  title  = {Qwen3-ASR-1.7B Korean OR Speech LoRA Adapter},
  author = {VitalDB / Seoul National University Hospital, Department of Anesthesiology and Pain Medicine},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/vitaldb/qwen3-asr-1.7b-kor-or-lora}
}

@dataset{synthetic_kor_speech_audio_v1_2026,
  title  = {Synthetic K-OR Speech Audio v1},
  author = {VitalDB / Seoul National University Hospital, Department of Anesthesiology and Pain Medicine},
  year   = {2026},
  publisher = {KHDP}
}

Acknowledgement

This work was supported by the Korea ARPA-H Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health and Welfare, Republic of Korea (Grant No. 2460006561).