Viterbi-guided method improves hotword selection for speech recognition
VISH-ASR: Viterbi-Guided Selection of Hotwords for Speech-LLM Contextual ASR
Artificial IntelligenceHuman-Computer Interaction
Summary
Speech recognition systems often need to find specific key words from large lists, but picking the right few quickly and accurately can be hard. The authors created a new method called VISH that checks how well each candidate word matches the audio using detailed information from the speech model, without requiring extra training. Their approach is faster and more accurate than older methods that either guessed from text or used separate models. This means speech systems can better recognize important words in noisy or complex environments with less computing power.
What this means in practice
- •For voice assistant developers: Improve voice assistant accuracy by quickly selecting relevant hotwords from large dynamic lists using acoustic verification without retraining models.$Commercial implications: Enables more accurate real-time hotword detection in voice assistants, enhancing customer experience while reducing computational costs.
- •For call center software teams: Enhance transcription tools by efficiently verifying keywords in spoken calls, reducing errors in dynamic vocabulary contexts.
Authors
Xiaoyu Gu, Yu Xi, Chencheng Wang, Zonghan Zhou, Haoyu Li, Jing Peng, Kai Yu
Abstract
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. 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. 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. 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. These results demonstrate the effectiveness of posterior-level verification for accurate and scalable hotword pre-selection without additional model training.