We brought together builders, researchers, and practitioners working on the cutting edge of voice AI for an evening of honest conversation. No pitches, no slides. Just the real challenges people are working through right now.

Four themes that defined the night

• Real-world audio quality

  Lab benchmarks don't tell the full story. Background noise, AirPods on a busy street, restaurant chatter. That's where speech models are actually tested, and closing that gap is still one of the hardest problems in the space.

• Latency is a full-stack problem

  Every layer adds up. Carrier hops, inference time, context window size. Owning your infrastructure changes what's possible, and the room had a lot to say about where the bottlenecks actually live.

• Inconsistent quality isn't just the model

  More often than not, inconsistent outcomes trace back to orchestration and configuration across the stack, not model performance itself. Tune your parameters. Benchmark on your own data.

• Dynamic context is unsolved

  Static keyword injection only goes so far. Getting voice agents to understand the full context of a call in real time, especially in high-stakes conversations, remains one of the biggest open problems in the space.

A huge thank you to Rohan, Anusha, and Ryan for sharing their insights throughout the evening, and to every attendee who spoke up and kept the conversation honest. Thanks to Telnyx for hosting and to AssemblyAI for co-sponsoring the night.

See you at the next one!