Behavior-aware
Evaluate pruning decisions through recovery behavior, rather than a single static proxy.
Efficient speech intelligence
Audio encoders,
made leaner.
X-AuT is a research project exploring behavior-driven progressive pruning for audio encoders in LLM-based automatic speech recognition.
Explore the project
01
Overview
Pruning guided by model behavior.
Modern speech-language systems rely on deep audio encoders. X-AuT studies how to reduce encoder depth while preserving the behaviors that matter for reliable recognition.
Progressive
Take controlled steps toward compact audio encoders.
Practical
Focus on efficient deployment for next-generation speech systems.
Coming soon