New AI Can Finally Understand Long Podcasts and Audiobooks
This AI can now listen to hours of audio and answer your questions about it — no manual notes needed.
AudioSpan is a new benchmark for audio comprehension that pushes both duration and depth, pairing audio clips from 10 minutes to over 2 hours with 3,240 questions across three cognitive levels: perception, understanding, and reasoning. The questions come from two automated paths—one that pulls questions directly from the audio content, and another that plants acoustic anchors to build a perception-to-reasoning chain scored to the first error. When the researchers evaluated 12 large audio-language models, they found the real challenge isn’t reasoning itself—it’s distilling a few relevant facts from a long, redundant signal. That difficulty grows as audio gets longer and hits hardest on perception, especially temporal grounding.
- AudioSpan is a new AI test that pushes voice AIs to understand hours-long audio, not just short clips
- Most current AI systems fail at pulling key details from long, noisy audio—especially when timing matters
- This could lead to better tools for summarizing meetings, analyzing training videos, and fact-checking interviews
Why It Matters
Soon, AI could accurately summarize your long meetings or podcasts—saving hours of your time.