Skip to content
Engineering

How we build voice AI for the phone.

The decisions, trade-offs, and physics behind PyAI - telephony audio, realtime latency, and the architecture under our speech models.

Engineering 6 min read

Why we built our speech model for 8 kHz — and why most voice AI fails on a real phone call

A phone line throws away half the audio. Models trained on studio sound demo great and crumble on real calls. Here's the physics, and what we did about it.

June 20, 2026Read

Ship phone agents on one speech-to-speech model.

Start with a free key, 10,000 Hear minutes every month, and $50 in credits. Build on the same OpenAI-compatible stack you take to production with Omni, Agents, and Trace - natural turn-taking, telephony-native, live in an afternoon.

No credit card - OpenAI-compatible - cancel anytime