OpenAI has engineered a new real-time voice system, GPT-Live, to significantly reduce audio latency in ChatGPT. By re-architecting the media pipeline and decoupling complex tasks, the system achieves a p95 audio frame delay that matches the old system's p50. Key innovations include a custom WARP protocol to speed up WebRTC connections and a dual-model approach that separates immediate conversational feedback from background tasks like search and tool use. This overhaul allows for continuous speech interaction, where the model can listen and respond simultaneously, managing conversational state and handling interruptions more effectively. AI
IMPACT Reduces latency in voice AI interactions, enabling more natural and fluid conversations.
RANK_REASON This is a detailed engineering breakdown of a product improvement, not a new model release or frontier research.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
