VISH-ASR: Viterbi-Guided Selection of Hotwords for Speech-LLM Contextual ASR

Key points

• Prompt-based contextual ASR requires selecting a compact hotword subset from large dynamic inventories, yet existing pre-selection methods rely either on discrete ASR hypotheses that may discard acoustic evidence or on separately trained retrieval models.1
• VISH formulates pre-selection as constrained posterior search; M-VISH makes large-list verification scalable through frame-synchronous shared-prefix Viterbi search with bounded top-K pruning, while Bi-MVISH exploits complementary backward evidence for robust verification.1
• Across five benchmarks, M-VISH achieves up to 11.7% absolute improvement in Best-F1 over a strong learned-retrieval baseline and reduces candidate-search latency by 75.2% at 2,000 candidates relative to hypothesis-based filtering.1
• These results demonstrate the effectiveness of posterior-level verification for accurate and scalable hotword pre-selection without additional model training.1

Sources (1)

• [1]VISH-ASR: Viterbi-Guided Selection of Hotwords for Speech-LLM Contextual ASR
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 08:49 AM

  We propose Viterbi-guided Selection of Hotwords (VISH), a training-free, plug-and-play framework that directly verifies candidate-specific acoustic support over frame-level CTC posteriors.

  Prompt-based contextual ASR requires selecting a compact hotword subset from large dynamic inventories, yet existing pre-selection methods rely either on discrete ASR hypotheses that may discard acoustic evidence or on separately trained retrieval models.

Extractive summary: sentences quoted from the sources.