--- title: How OpenAI rebuilt its voice stack for GPT-Live and two... description: OpenAI released two new transcription models — GPT-Live-Transcribe for low-latency live transcription and GPT-Transcribe for async/batch workloads — alongside... canonical: https://daily.dev/posts/how-openai-rebuilt-its-voice-stack-for-gpt-live-and-two-new-transcription-models-uvjluticb twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How OpenAI rebuilt its voice stack for GPT-Live and two new transcription models | daily.dev og:description: OpenAI released two new transcription models — GPT-Live-Transcribe for low-latency live transcription and GPT-Transcribe for async/batch workloads — alongside... og:url: https://daily.dev/posts/how-openai-rebuilt-its-voice-stack-for-gpt-live-and-two-new-transcription-models-uvjluticb og:image: https://api.daily.dev/og/posts/uvjLuTIcB.png og:image:alt: How OpenAI rebuilt its voice stack for GPT-Live and two new transcription models og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How OpenAI rebuilt its voice stack for GPT-Live and two new transcription models **[Collections](https://daily.dev/sources/collections)** · 3 min read · 1 upvotes · 0 comments ## Summary OpenAI released two new transcription models — GPT-Live-Transcribe for low-latency live transcription and GPT-Transcribe for async/batch workloads — alongside architectural details of GPT-Live, their third-generation voice system. The key shift is moving from a turn-detector model to full-duplex audio, allowing simultaneous listening and speaking. The media frontend was rewritten in Go (replacing Python asyncio) for smoother frame delivery. Transport uses WebRTC with a custom WARP protocol that cuts session handshake from six round trips to one, plus an Instant Connect mechanism. The system features stateful inference with seamless model handoffs, dynamic context compaction, and async delegation to frontier models like GPT-4.5 for heavy reasoning tasks. Validation used shadow testing on real ChatGPT Voice traffic rather than synthetic load tests. The entire system went from concept to production in six months. ## Content OpenAI recently published a detailed breakdown of how it built GPT-Live, a third-generation voice AI system, in six months. The result is a significant architectural departure from how voice AI has traditionally worked. ## Continuous audio instead of turn detection Previous voice systems relied on a separate turn detector to decide when one speaker had finished and the other could begin. GPT-Live drops that model entirely. It listens and speaks simultaneously, full-duplex, which means the system can pick up on interruptions naturally rather than waiting for a signal that a turn has ended. Deeper reasoning and tool calls run in parallel in the background, so a slow API call or complex query doesn't freeze the conversation. The user keeps hearing audio while a stronger model works on the response. ## Separating media transport from application logic One of the core engineering decisions was decoupling the media fast path from application logic. In practice, this means a slow tool call can delay its result without stalling the audio stream. The conversation keeps flowing regardless of what's happening in the background. The team also rewrote the media frontend and inference logic in Go, replacing Python asyncio. The reported improvement: the new system's p95 frame delivery matches the old system's p50. That's a meaningful jump in consistency at the tail end of the latency distribution, where voice conversations feel choppy. ## Session startup and WARP WebRTC is the transport layer, but standard WebRTC setup requires six network round trips before a session begins. OpenAI built WARP (WebRTC Abridged Roundtrip Protocol) to cut that down to one. They also added an "Instant Connect" mechanism that can establish a session from a single UDP packet, making the moment a user starts a voice conversation feel nearly instantaneous. ## Handling long calls and context growth Long voice sessions create two problems: model instances eventually need to be replaced, and context windows fill up. GPT-Live handles both with the same mechanism. A replacement model is prefilled from the current conversation state before traffic switches over, so the handoff is seamless. The old model keeps talking until the new one is ready. The same approach handles context compaction: the conversation continues uninterrupted while the context is trimmed in the background. ## Production validation The team validated the new system through shadow testing on real ChatGPT Voice traffic rather than synthetic load tests. That turned out to matter. Real traffic exposed capacity, geography, and observability issues that wouldn't have shown up in a controlled environment. ## New transcription models in the API Alongside GPT-Live, OpenAI also released two new transcription models: - **GPT-Live-Transcribe**: built for low-latency live transcription - **GPT-Transcribe**: optimized for asynchronous transcription of completed audio files and batch workloads Both models are designed to handle real-world audio better than previous options, including short phrases, numbers, specialized terminology, heavy accents, and loud background noise. ## Questions this post answers ### How does GPT-Live handle turn-taking differently from previous voice AI systems? GPT-Live eliminates the separate turn detector model used by prior voice AI systems and instead listens and speaks simultaneously in full duplex. This lets it naturally handle interruptions instead of waiting for a signal that indicates one speaker has finished, while deeper reasoning and tool calls run in parallel in the background without freezing the conversation. _daily.dev surfaces engineering deep dives like this for developers architecting real-time voice systems._ ### What is WARP and how does it speed up WebRTC session startup? WARP, short for WebRTC Abridged Roundtrip Protocol, is a mechanism built by OpenAI to reduce WebRTC session startup from six network round trips down to one. It also pairs with an Instant Connect feature that can establish a voice session from a single UDP packet, making conversation start feel nearly instantaneous. _Developers optimizing real-time transport latency can find protocol breakdowns like this on daily.dev._ ### Why did OpenAI rewrite the GPT-Live media frontend in Go instead of Python asyncio? OpenAI rewrote the media frontend and inference logic in Go to improve tail-latency consistency, reporting that the new system's p95 frame delivery now matches the old Python asyncio system's p50. This targets the choppiness voice conversations experience at the tail end of the latency distribution. _Engineers weighing Go versus Python for latency-sensitive systems track cases like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion across x (as of 2026-09-13). **TL;DR:** There is no community discussion available to analyze for this post — no comments or replies were provided. **Sentiment:** 50% positive · 30% mixed · 20% skeptical **By community** - x (mixed): No replies were available to assess sentiment. **Source threads** - [x](https://x.com/rohanpaul_ai/status/2084666052626235758) · 0 points · 0 comments ## Similar posts on daily.dev - [OpenAI Details GPT-Live’s Architecture for Continuous Stateful Voice Interaction](https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop) · InfoQ · 1 upvotes · 0 comments - [OpenAI's GPT-Live: ChatGPT voice that listens and talks](https://daily.dev/posts/openai-s-gpt-live-chatgpt-voice-that-listens-and-talks-nwap9djkb) · The Next Web · 9 upvotes · 3 comments --- Tags: [#golang](https://daily.dev/tags/golang), [#openai](https://daily.dev/tags/openai), [#speech-recognition](https://daily.dev/tags/speech-recognition), [#webrtc](https://daily.dev/tags/webrtc) [View this post on daily.dev](https://daily.dev/posts/how-openai-rebuilt-its-voice-stack-for-gpt-live-and-two-new-transcription-models-uvjluticb) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How OpenAI rebuilt its voice stack for GPT-Live and two new transcription models","url":"https://daily.dev/posts/how-openai-rebuilt-its-voice-stack-for-gpt-live-and-two-new-transcription-models-uvjluticb","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-openai-rebuilt-its-voice-stack-for-gpt-live-and-two-new-transcription-models-uvjluticb"},"datePublished":"2026-08-03T22:26:18.348Z","dateModified":"2026-09-13T19:20:30.769Z","description":"OpenAI released two new transcription models — GPT-Live-Transcribe for low-latency live transcription and GPT-Transcribe for async/batch workloads — alongside...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-openai-rebuilt-its-voice-stack-for-gpt-live-and-two-new-transcription-models-uvjluticb","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"golang,openai,speech-recognition,webrtc","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"How OpenAI rebuilt its voice stack for GPT-Live and two new transcription models"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-openai-rebuilt-its-voice-stack-for-gpt-live-and-two-new-transcription-models-uvjluticb#faq","mainEntity":[{"@type":"Question","name":"How does GPT-Live handle turn-taking differently from previous voice AI systems?","acceptedAnswer":{"@type":"Answer","text":"GPT-Live eliminates the separate turn detector model used by prior voice AI systems and instead listens and speaks simultaneously in full duplex. This lets it naturally handle interruptions instead of waiting for a signal that indicates one speaker has finished, while deeper reasoning and tool calls run in parallel in the background without freezing the conversation. daily.dev surfaces engineering deep dives like this for developers architecting real-time voice systems."}},{"@type":"Question","name":"What is WARP and how does it speed up WebRTC session startup?","acceptedAnswer":{"@type":"Answer","text":"WARP, short for WebRTC Abridged Roundtrip Protocol, is a mechanism built by OpenAI to reduce WebRTC session startup from six network round trips down to one. It also pairs with an Instant Connect feature that can establish a voice session from a single UDP packet, making conversation start feel nearly instantaneous. Developers optimizing real-time transport latency can find protocol breakdowns like this on daily.dev."}},{"@type":"Question","name":"Why did OpenAI rewrite the GPT-Live media frontend in Go instead of Python asyncio?","acceptedAnswer":{"@type":"Answer","text":"OpenAI rewrote the media frontend and inference logic in Go to improve tail-latency consistency, reporting that the new system's p95 frame delivery now matches the old Python asyncio system's p50. This targets the choppiness voice conversations experience at the tail end of the latency distribution. Engineers weighing Go versus Python for latency-sensitive systems track cases like this on daily.dev."}}]} ```