X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
X-AuT cuts speech-model audio-encoder layers while keeping error rates near, or in one case below, the full model.
The paper compresses Qwen3-ASR-0.6B by pruning audio-encoder depth, then repairing the model with alignment, cross-scale distillation, scheduled supervision, and LoRA finetuning. Its 16-layer version reports a macro-average error of 5.27%, down from 5.61% for the 18-layer baseline across ten Chinese-English benchmarks. A 14-layer version reaches 5.75% with 20.7% fewer audio-tower parameters. The authors say progressive pruning beats direct pruning, and that benchmark-level accuracy shifts vary. HF Daily Papers' note
The paper compresses Qwen3-ASR-0.6B by pruning audio-encoder depth, then repairing the model with alignment, cross-scale distillation, scheduled supervision, and LoRA finetuning. Its 16-layer version reports a macro-average error of 5.27%, down from 5.61% for the 18-layer baseline across ten Chinese-English benchmarks. A 14-layer version reaches 5.75% with 20.7% fewer audio-tower parameters. The authors say progressive pruning beats direct pruning, and that benchmark-level accuracy shifts vary. HF Daily Papers' note
score 5