Reproduce “Think Before You Transcribe: Context-Steered Reasoning for Hotword-Aware Speech Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qihao Wang, Hang Su, Fan Cui, Lichun Fan, Jian Luan, Yue Hu
CAST asks an audio recognizer to produce a grounded caption before transcribing, using retrieved candidate names only when they appear supported by the audio. On RACE-Bench, a 30B Qwen3-Omni model reaches 3.2% WER and 12.9% NE-FNR on retrieved-context averages, versus 3.6% and 17.2% for the baseline. In all-noise tests, WER is 6.3% versus 6.2% contextless; evidence is limited to English–Chinese benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.