Prompt-based contextual speech recognition needs to choose a small set of hotwords from inventories that can be large and change frequently, because those items are placed directly into the model's prompt. The new arXiv paper "VISH-ASR" argues that the way this pre-selection happens matters: methods built on discrete ASR hypotheses may throw away acoustic evidence before the full model gets a chance to use it.
The authors propose a Viterbi-guided selection method for speech-LLM systems. As the name suggests, Viterbi decoding guides which hotwords are kept from the inventory, with the goal of making prompt-based decoding more effective. The provided abstract is truncated, so the paper's full comparison against alternative pre-selection strategies is not yet visible; the only contrast explicitly stated is with methods that rely on discrete hypotheses.
The contribution is framed as a selection mechanism for large dynamic inventories rather than a change to the recognizer or language model itself. No experimental numbers or baselines are available from this source alone.