Kadri Hacioglu

A discriminative adaptation that reads the final prompt token's hidden state through a classification head, producing a label in one forward pass without modifying the backbone, gives a controlled comparison of generative and discriminative inference in an otherwise identical speechLLM.

These findings suggest that turn-taking is primarily conveyed through intonation and silence patterns rather than semantic completeness, enabling faster and more reliable systems without expensive text inference.

Just 4 hours of speech data closes the modality gap in LLM-based ASR, rivaling full-dataset fine-tuning and unlocking effective domain adaptation.

LLM-based ASR can get a context boost without the compute cost: compress prior audio turns into learned latent tokens and retain transcripts to recover accuracy while shrinking the audio footprint.