SURE-Voice: A Front-End Baseline for Speech-Evidence Filtering in Speech LLMs
Mengzhe Geng's paper on SURE-Voice addresses how speech LLMs handle irrelevant audio like noise or silence. The discussion covers the SURE-Challenge, current model failures in rejecting unsupported inputs, and the effectiveness of using lightweight front-end filters, such as Whisper-score thresholds, to improve efficiency and reliability.