{"rewrite":{"id":"r_23bda6a56245aec8dab8deba","clusterId":"c_313e2fed4a2e0794a82f7daf","slug":"openai-explains-how-gpt-live-cuts-voice-ai-response-latency","model":"deepseek-v4-flash:free","headline":"OpenAI Explains How GPT-Live Cuts Voice AI Response Latency","summary":"OpenAI detailed the realtime system behind GPT-Live, its voice AI for natural conversation. The system uses full-duplex communication, offloads complex tasks to other models, and a new protocol called WARP that reduces network round trips from six to one. GPT-Live powers ChatGPT Voice, with a future API planned.","whyItMatters":"OpenAI's explanation shows how GPT-Live avoids the unnatural pauses and interruptions of earlier voice AI by letting the model talk and listen simultaneously, a design that could shape future voice interfaces.","webCardHtml":"\u003cp\u003eOpenAI has published a technical breakdown of GPT-Live, the voice AI that powers ChatGPT Voice, explaining how it responds without the awkward silences or interruptions common in earlier systems. The key change is moving speech-end detection out of the audio pipeline and adopting full-duplex communication, where both sides send and receive audio at once, like a phone call. GPT-Live decides to speak, keep listening, wait, or interrupt many times per second.\u003c/p\u003e\u003cp\u003eFor tasks requiring search or complex reasoning, GPT-Live hands the work to a more capable model while conversation continues over a dedicated fast path. Long sessions prepare replacement model instances in the background, loading context before switching so users do not notice the change. The audio delivery uses WebRTC, but OpenAI developed WARP to cut the network round trips needed to start a session from six to one. OpenAI says it plans to offer a GPT-Live API to spread voice interaction to more devices and apps.\u003c/p\u003e","blueskyPost":"OpenAI explains how GPT-Live keeps voice chats natural: full-duplex audio, offloading hard tasks, and WARP cutting setup round trips from 6 to 1. It powers ChatGPT Voice, with an API coming.","twitterPost":"OpenAI details GPT-Live's realtime voice system: full-duplex talk/listen, task offloading, and WARP cutting network round trips from 6 to 1. It runs ChatGPT Voice, API planned.","threadsPost":null,"newsletterBlurb":"OpenAI published how GPT-Live achieves responsive voice AI: full-duplex communication, offloading complex tasks, and a new WARP protocol that cuts setup round trips from six to one. The system powers ChatGPT Voice, with a future API to expand voice interaction.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260804-openai-gpt-live/\",\"title\":\"音声AIはなぜすぐに返事できるのか？OpenAIが「GPT-Live」の仕組みを解説\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4933,"outputTokens":595,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1786321789,"createdAt":"2026-08-10T00:18:32.000Z","publishedAt":"2026-08-10T00:21:44.000Z","updatedAt":"2026-08-10T00:18:32.000Z"},"cluster":{"id":"c_313e2fed4a2e0794a82f7daf","canonicalTitle":"音声AIはなぜすぐに返事できるのか？OpenAIが「GPT-Live」の仕組みを解説","representativeArticleId":"a_fc3e75ee58b565fa4502645c","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"GPT-Live\"],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-08-04T06:35:00.000Z","lastSeenAt":"2026-08-04T06:35:00.000Z","updatedAt":"2026-08-10T00:21:45.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260804-openai-gpt-live/","title":"音声AIはなぜすぐに返事できるのか？OpenAIが「GPT-Live」の仕組みを解説"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["GPT-Live"],"studios":[],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":null}