How to Recognize New Words: A Comparison Between Context Biasing Methods and Speech LLMs

The study evaluates two context biasing approaches against speech LLMs prompted with context, focusing on named entities, acronyms, and domain-specific terms that are scarce in training data. Context biasing extends an ASR model to accept a word list during inference, while speech LLMs use direct prompting. The paper likely provides empirical comparisons on benchmark datasets, though specific results are not in the excerpt. This work is significant because rare word recognition is critical for real-world ASR applications like voice assistants and dictation. Future developments may see hybrid approaches combining context biasing with LLM prompting.