Beyond Prompting: Efficient and Robust Contextual Biasing for Speech LLMs via Logit-Space Integration (LOGIC)

LOGIC integrates contextual entity lists directly into the logit space of Speech LLMs to improve domain-specific recognition without increasing context window length. This method avoids the inference latency and hallucination risks associated with prompting or generative error correction. It enables real-time adaptation to personalized or evolving technical jargon.