GPT-Live-1 Meets Your Own Backend: Turn-Taking, Silence-Filling, and Who Owns the Call
Slight_Republic_4242 · reddit · 2026-09-15
Dograh's founder breaks down wiring GPT-Live-1 as a voice layer with any backend: the benchmark score was measured with GPT-6 Astra behind it, so your number depends on your own stack. Two hard-won lessons: (1) the voice and reasoning models don't share a clock — GPT-Live-1 is trained to never sit in silence, and on a customer call, filling is committing; (2) you hand turn-taking to the vendor — OpenAI owns the VAD signals, Gemini Live splits it differently, Ultravox hands it back. Speech-to-speech also sacrifices the cascaded stack's observability, and one vendor now owns the whole call while training on your data and serving competitors. 'The key takes an afternoon. The system around the key is the product.'
Related event: GPT-Live-1 Benchmark Scores Depend on Backend Model(2 posts)→
More from coding & agent
- Tencent unveils Hyra 2.0, a research agent that builds reusable scientific tools — TencentHunyuan · 2026-11-02
- What broke when I ran a leaderless agent swarm with no orchestrator — 7 failure modes and fixes — LostHumanity67 · 2026-11-02
- Forced yes/no prompts push GPT and Claude to 95%+ confidence on Goldbach's unproved halt test — Nir777 · 2026-11-02
- Automation took over my daily work — so why am I still sitting in the office for 8 hours? — Am1rSori · 2026-11-02
- General Reasoning explains how to train agent swarms with RL and latency pressure — rosstaylor90 · 2026-11-02
- "Coding isn't the bottleneck anymore — usage windows are," says founder Kyle Gawley — kylegawley · 2026-11-02