LoRA adapter (the trained delta, ~95 MB) for Qwen/Qwen3-ASR-1.7B, adapting it to degraded VHF airband (ATC) radio. Applied to both the audio encoder and the text decoder (all-linear, r=16, α=32).

Most users want the pre-merged, drop-in model instead: chrullis/qwen3-asr-radio-1.7b (self-contained, serves in vLLM directly). This adapter is for composing/inspecting the delta or applying it yourself.

Results (WER, lower is better)

eval set	base	+ this adapter
UWB-ATCC test (held-out)	0.766	0.317
ATCO2 (cross-domain, unseen speakers)	0.609	0.327
LibriSpeech (clean, forgetting check)	0.013	0.010

Applying the adapter

Qwen3-ASR loads via the qwen-asr package (pip install qwen-asr peft); the adapter attaches to the model's Thinker (encoder + decoder), which needs a small get_input_embeddings shim for PEFT:

python

import types, torch, torch.nn as nn
from peft import PeftModel
from qwen_asr.core.transformers_backend.modeling_qwen3_asr import Qwen3ASRForConditionalGeneration

top = Qwen3ASRForConditionalGeneration.from_pretrained("Qwen/Qwen3-ASR-1.7B", dtype=torch.bfloat16)
emb = max((m for m in top.thinker.modules() if isinstance(m, nn.Embedding)), key=lambda e: e.num_embeddings)
top.thinker.get_input_embeddings = types.MethodType(lambda self: emb, top.thinker)
top.thinker = PeftModel.from_pretrained(top.thinker, "chrullis/qwen3-asr-radio-1.7b-lora").merge_and_unload()
# `top` is now the fine-tuned model (equivalent to the merged repo).

Serving with vLLM

Use the pre-merged repo chrullis/qwen3-asr-radio-1.7b (it has the vLLM instructions). To serve this adapter, merge it first — extend the snippet above with top.save_pretrained("merged"); proc.save_pretrained("merged"), then serve merged/ in vLLM like any Qwen3-ASR model.

Do not try vLLM runtime LoRA (--enable-lora) here: this adapter also trains the audio encoder, which vLLM's LoRA serving doesn't load (it targets LLM/decoder layers), and qwen3_asr runtime-LoRA support is unverified. Merging is the correct path.

Intended use, limitations, training, licensing

Same as the merged model — see chrullis/qwen3-asr-radio-1.7b. In brief: assistive ATC-airband transcription (not unattended verbatim); aviation-English only; trained on real UWB-ATCC + synthetic-degraded LibriSpeech/ATCOSIM with EOS-terminated targets; verify the academic-corpus terms before commercial redistribution.