ASR models lack native semantic prompting support

The absence of semantic prompting in ASR limits the effectiveness of voice agents in specialized domains like license plate recognition or medical terminology.

• –Current "word boosting" techniques are brittle and don't scale to broad categories or long context.
• –Fine-tuning models to accept <text> prompts could allow for zero-shot boosting of specific semantic classes (e.g., "Australian cities").
• –Feeding conversation history directly into the ASR layer could significantly improve transcript accuracy for multi-turn voice interactions.
• –Implementation likely lags due to training data scarcity for prompted ASR and the computational overhead of cross-modal context.