🎙️ X-AuT — Compressed Speech-LLM ASR

X-AuT progressively compresses the audio encoder of Qwen3-ASR-0.6B from 18 to 14 transformer blocks (−20.7% audio-tower parameters) and recovers the accuracy with cross-scale distillation from a frozen Qwen3-ASR-1.7B teacher. Audio-in → transcript-out.

Upload or record audio (any common format, resampled to 16 kHz mono) and press Transcribe.

Notes. Example audio: LibriSpeech test-clean utterances (CC-BY-4.0) and AISHELL-1 utterances (Apache-2.0). The demo runs the authors' inference recipe (infer_xaut.py) on ZeroGPU. Model weights are CC BY-NC 4.0 — research / evaluation use only.