Four voice AI talks from the AI Engineer World's Fair. All under 15 minutes. 36 minutes total. That's less than your commute, and more useful than most of what's in your feed this week. Here's what to watch, in the order I'd watch them. 1. Your Voice Agent Doesn't Need a Frontier Model Joel Allou and Ornella Bahidika, Microsoft. 6 min. The one to start with. Most teams reach for the biggest model by default. In voice, every millisecond of that choice is dead air on a phone line. 2. Don't Let the LLM Drive Same speakers, Microsoft. 6 min. Watch it straight after the first. Same team, the other half of the argument: what the model should decide, and what it never should. Twelve minutes in, you'll already have changed how you think about your stack. 3. Voice In, Visuals Out: The Agony and the Ecstasy Allen Pike. 13 min. Voice agents that also show you something. The title promises both the good and the painful parts, which is exactly what you want from a talk like this. 4. Every Solo Agent Builder Eventually Reinvents a Worse Version of CI/CD Sumaiya Shrabony. 11 min. Not voice-specific. Watch it anyway. If you ship agents alone, you are already building this badly. Better to hear it now than discover it at 2am. Why these four together: Two on what the model should and shouldn't do. One on what voice looks like beyond audio. One on how you ship without breaking things. That's the whole job, in 36 minutes. Full list, filtered to speech and audio, under 15 min: ai.engineer/talks?topic=spee… Which one are you starting with?