Distilling Conversations: Abstract Compression of Conversational Audio Context for LLM-based ASR

Dimensions overall score 4.0

GitHub Code Pulse

No public code linked for this paper yet.

• Conversational context, after supervised multi-turn training, primarily improves the recognition of contextual entities in LLM-based ASR.result 90%
• Abstract Compression replaces the audio portion of prior turns with a fixed number of learned latent tokens while retaining corresponding transcripts explicitly.method 95%
• The compressed model (Abstract Compression) recovers part of the performance gains of raw-context conditioning with a smaller prior-turn audio footprint.result 90%
• Conditioning on raw conversational context in LLM-based ASR is computationally expensive due to the rapid growth of the prior-turn audio token sequence with conversation length.limitation 90%
• Abstract Compression was trained by optimizing a compression module (turn-specific queries and cross-attention parameters) while keeping the base LLM frozen.method 90%
• On the WoW dataset, Abstract Compression reduced Word Error Rate (WER) from 25.6% to 23.3%.result 85%
• Retaining prior-turn transcripts explicitly in Abstract Compression is done because compressing text did not yield comparably effective results in preliminary experiments.method 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card