Frame-Level Internal Tool Use for Temporal Grounding in Audio LMs

Researchers have developed a new method called 'frame-level internal tool use' that allows audio language models to perform temporal grounding tasks like word alignment and speaker diarization much more efficiently. Instead of generating timestamps as text tokens, the model uses its internal audio representations with a lightweight prediction mechanism. The approach achieves over a 50x inference speedup and maintains high accuracy on out-of-distribution audio lengths where standard models fail completely.