Pıer
TidesCurrentsHarbor LightsLabBottlesAshore
Pıer

Navigation

  • Tides
  • Ashore
  • Harbor Lights
  • Agent Access
  • Changelog
  • Bottles
  • Now
  • Feedback

External links

GitHubCloudborne ↗

© 2026 Pier.

Read original
HuggingFace Daily Papers·Haojun Zhang·Sep 9, 2026, 8:00 PM

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs via Cross-Scale Distillation

Original title:X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation

Papers76

Trimming audio-encoder depth cuts inference overhead in speech LLMs, but directly dropping blocks often perturbs decoder embeddings, triggering token deletions and premature end-of-sequence stops. X-AuT addresses this with a progressive compression framework that selects layer subsets through behavioral probes and recovers performance via cross-scale distillation and LoRA tuning while keeping the language model backbone frozen. Evaluated on Qwen3-ASR-0.6B across ten Chinese-English benchmarks, pruning the audio encoder from 18 to 14 layers reduced audio-tower parameters by 20.7% with a macro-average error of 5.75%, outpacing direct pruning.

Why it's worth reading

As speech LLMs shift toward production serving, X-AuT presents a practical recipe to compress audio towers by over 20% while preventing decoding errors caused by naive layer pruning.

Tags

Speech LLMModel CompressionAudio EncoderKnowledge DistillationPruningQwen-ASREfficiency

Score breakdown

  • Novelty75
  • Impact73
  • Practicality82
  • Credibility76
  • Timeliness74