# OpenAI GPT-Live drops the transcript, and that is the whole point

> What is GPT-Live: OpenAI new native voice model claims sub 300ms latency in ChatGPT Voice by deleting the speech to text step in the middle of the pipeline.

- Published: 2026-08-31
- Topic: [AI](https://futuretechnologyhq.com/topic/ai/)
- URL: https://futuretechnologyhq.com/article/openai-gpt-live-voice-model/
- Publisher: [Future Technology](https://futuretechnologyhq.com/), free to read. Quote as "Future Technology" with a link to the article.

## Key points

- GPT-Live is a native voice model, so audio stays as audio instead of being flattened into a transcript
- OpenAI claims sub 300 millisecond latency, roughly the gap between turns in natural human conversation
- These are launch figures, and real latency over a mobile network will be worse

## Article

Three hundred milliseconds is roughly the gap between turns in ordinary human conversation. That is the number OpenAI is claiming for GPT-Live, the native voice model now running behind ChatGPT Voice.

The claim rests on removing a step rather than adding one. Voice assistants have generally worked in three stages: speech to text, then the language model, then text to speech. Every stage adds delay, and every stage throws information away. Tone, hesitation, emphasis, the half second where somebody changes their mind mid-sentence, all of it dies in the transcript.

## What is GPT-Live?

GPT-Live is OpenAI's native voice model, and it now powers ChatGPT Voice. "Native" means the model takes in sound and produces sound directly, rather than handing work between three separate systems. According to the launch coverage from AI Weekly and the Artificial Intelligence Herald, the headline claim is a latency under 300 milliseconds, which is the delay between you finishing a sentence and the model beginning its reply.

Latency is the figure that decides whether a voice assistant feels usable. A pause of a second or two is tolerable for a search box, but in speech it reads as confusion or boredom. People notice a late reply almost immediately, even if they could not say why the conversation feels off.

## How did voice assistants work until now?

The older design is called a cascaded pipeline. A speech recognition model listens and writes out what it thinks you said. A language model reads that text and writes a reply. A third model turns the reply into a synthetic voice. Each handoff has to wait for the previous stage to finish, or at least to produce enough output to start on.

The delay is only part of the problem. A transcript is a lossy record. It keeps the words and drops nearly everything else: whether you sounded sure, whether you were laughing, whether you stopped because you were interrupted. The language model in the middle only ever sees the flattened version, so it cannot respond to what was never passed to it.

Assistants built this way have shaped how people use voice since 2011, when the first mainstream phone assistants arrived. Most users learned to speak in short, clean commands and to wait politely for the answer.

## What native voice changes

A native voice model keeps the audio as audio the whole way through. There is no transcript in the middle to reconstruct meaning from, so the model can hear that a question was tentative rather than just reading the words of it, and can answer with matching emphasis.

The latency figure matters more than the emotional nuance, at least in the short term. Below roughly 300 milliseconds, interruption starts working properly. You can cut in, the model stops, and the exchange feels like talking rather than taking turns with a kiosk. Above it, everyone falls back into the walkie-talkie rhythm that has made voice assistants tiring since 2011.

Interruption is the practical test. In a cascaded system, the recogniser must decide you have started speaking, cancel the reply that is playing, and restart the chain. A model that processes sound continuously can treat your voice as just another input arriving mid-answer.

## How does it compare with the older approach?

| | Cascaded pipeline | Native voice (GPT-Live) |
|---|---|---|
| Stages | Speech to text, language model, text to speech | One model handling audio in and audio out |
| What the model receives | A text transcript | The audio itself |
| Tone and hesitation | Lost at the transcription step | Available to the model |
| Interruption | Awkward, needs the chain to restart | Designed to work below roughly 300 milliseconds |
| Text record for logging | Produced as a by-product | Has to be generated separately |

The last row is the one that gets least attention, and it comes back below.

## The caveat worth stating plainly

These are launch claims measured under launch conditions. Real latency over a mobile network, on a congested evening, with a cheap Bluetooth headset in the chain, will be higher. Bluetooth alone can add well over 100 milliseconds before the model has heard anything.

That matters because the budget is tight. If 300 milliseconds is the target and a wireless headset takes a third of it before the audio even leaves your pocket, the model has to be very quick to land anywhere near the claim. Wired headphones and a good connection will get closer to the advertised experience than the average commute will.

Where the processing happens matters too, and that trade is the same one covered in our comparison of [local AI agents against cloud agents](/article/local-ai-agent-vs-cloud-agent/). Native voice is heavy enough that it stays in the cloud for now, which means the network is part of the latency budget whether OpenAI likes it or not.

## Who else is working on voice assistants?

OpenAI is not alone in rebuilding voice. Apple has been reworking its assistant too, and we covered how [Apple shipped the new Siri with Google's model doing the talking](/article/apple-siri-gemini-public-beta-ios-27/). The same release lets people adjust how the assistant sounds, which 9to5Mac and TechCrunch reported when the iOS 27 beta 3 sliders went live, and which we looked at in [the iOS 27 beta controls for Siri's pace and expressiveness](/ios-27-siri-pace-expressivity-beta/).

The common thread is that the voice itself is now treated as part of the product, not a thin layer over a text chatbot. How a reply is delivered, how quickly, and how naturally it handles being cut off are becoming the points of competition, alongside how clever the answers are.

## What does this break for businesses?

The interesting question is what happens to everything built on the assumption that a transcript exists. Logging, moderation, analytics and most agent tooling all read text. If the audio never becomes text, that plumbing has to be rebuilt or run as a separate pass, which is a cost nobody mentions in a launch post.

A company that records customer calls for compliance, for example, wants a searchable written record. With a native model, producing that record becomes an extra job instead of a free side effect. Agent interoperability work such as [A2A and MCP](/article/a2a-protocol-vs-mcp/) is written around text payloads as well, so voice agents will need a translation layer before they slot into those systems.

None of this is a reason the approach is wrong. It is a reminder that removing the transcript removes a convenient inspection point, and someone has to decide what replaces it.

## What to watch

Watch for independent latency measurements over the next few weeks. That number, taken on a normal connection rather than a demo stage, is the one that decides whether this changes how people use voice or just how it sounds.

Also watch how OpenAI handles records and moderation for audio, since that is where the practical friction will show up first. If those answers arrive quietly and the latency holds up, native voice stops being a launch feature and becomes the default way these assistants are built.

## Key takeaways

- OpenAI claims that GPT-Live, the native voice model behind ChatGPT Voice, has a latency of around 300 milliseconds.
- A native voice model keeps audio as audio throughout the process, without creating a transcript in the middle.
- Below 300 milliseconds, interruption in conversation starts to work properly, making the exchange feel like talking rather than taking turns.
- Real latency may be higher than the launch claims, because Bluetooth alone can add well over 100 milliseconds.
- Without a transcript, logging, moderation and analytics built for text have to be rebuilt or run as a separate pass.

## Sources

- [AI Weekly](https://aiweekly.co/ai-news-today)
- [Artificial Intelligence Herald](https://artificialintelligenceherald.com/ai-news-today)