ReListen Enables Training-Free Evidence Localization in Long Audio Files
November 4, 2026
ReListen allows frozen audio language models to locate specific evidence in hour-long recordings without retraining. By using a Jev-style scan to score 20-second windows, the method achieves an 88.5% overlap with annotated evidence on VoiceGiraffe questions, outperforming transcript-based retrieval.
HOW THIS AFFECTS YOU
●
builderYou can now implement long-form audio QA using existing frozen models without expensive fine-tuning.
●
researcherThe method demonstrates that probability-based window scoring can replace explicit temporal grounding training.