Abstract Compression of Conversational Audio Context for LLM-based ASR
Researchers investigated how conversational context from previous turns improves Large Language Model (LLM) speech recognition (ASR). They found that multi-turn training helps, especially for recognizing contextual entities. To make using raw context cheaper, they introduced Abstract Compression, which replaces long audio sequences with fixed latent tokens while keeping transcripts. This method recovers significant performance gains with a much smaller context footprint.