LiveKit: The Open-Source Platform Behind OpenAI's Voice Mode
About this episode
Tom Shapland's platform powers ChatGPT's voice mode and handles over 100,000 minutes of calls a day from a single customer. He walks through what breaks when you build a voice agent the naive way, explains the turn detection system that determines whether AI sounds human or robotic, and confesses why he'll never do B2B sales again despite building a unicorn.
Key moments
- 0:00 intro
- 1:49 Voice AI explained to an 8-year-old
- 4:22 Building a voice agent the naive way
- 7:44 Why your voice agent responds so slowly
- 10:07 Speech-to-speech vs pipeline in production
- 14:32 Why ChatGPT keeps interrupting you
- 19:34 Managing five different API keys
- 22:22 Coding agents build broken voice apps
- 24:42 Why people talk more to fake humans
- 32:30 100,000 voice minutes per day
- 34:08 Useful" startups fail, "weird" ones win
- 42:07 $1B moat anyone can fork
- 44:18 AI replacing junior engineers?
- 48:47 Why he'll never do sales again
- 56:46 Can a machine have judgment?
- 1:00:05 Rapid fire + Homer's Odyssey