OpenAI has announced GPT-Live-1.1, which adds new voices for French, German, Hindi, Mexican Spanish, Korean, and Japanese, improves accents, and reduces repetitive “checking” announcements. The update begins rolling out on 15 October 2026 to Voice in ChatGPT, Codex and dots. OpenAI says GPT-Live-1.1 will also come to the API, but has not provided a date.
For creators who need downloadable narration rather than a live conversation, DIY AI Studio’s AI voice generator converts written scripts into spoken audio. The OpenAI release addresses a different problem: whether a voice assistant sounds natural while responding, changing languages and handing off work.
What changes with GPT-Live-1.1?
| Update | Announced change | What remains unconfirmed |
|---|---|---|
| New voices | French, German, Hindi, Mexican Spanish, Korean and Japanese | Exact voice choices available to each account |
| Better accents | Improved speech quality across the new voice options | Performance for individual regional accents and dialects |
| Less “checking” | Fewer repetitive spoken status messages | Whether underlying search or reasoning takes less time |
OpenAI did not provide language-by-language accuracy scores, new latency measurements, or a full API specification for version 1.1 in the announcement. Those details don’t follow from the new model name.
Fewer “checking” messages address a real conversational annoyance
One recurring complaint about GPT-Live is its habit of announcing that it is “checking” or “looking something up” even during ordinary conversation. After hearing the same phrase repeatedly, some people find it harder to maintain a natural rhythm. Others prefer a spoken acknowledgement to a long, unexplained silence. The useful change, therefore, is better judgement about when an update is necessary, not simply removing every progress message.
Fewer spoken checks do not necessarily mean fewer web searches or faster reasoning. GPT-Live can delegate a difficult request to a backend model or tool while maintaining the spoken exchange. GPT-Live-1.1 may make those handovers less intrusive without changing the amount of work required. A natural voice should still communicate uncertainty when the facts need verification, particularly for current information or actions that could fail.
Why Mexican Spanish is a particularly useful accent test
Mexican Spanish is a more specific promise than “improved Spanish”. A speaker in Mexico may expect familiar vocabulary, natural local place-name pronunciation and a recognisable rhythm, not an accent aimed at Spain or a generic Latin American delivery. Each of the six announced language options deserves comparable checks by people familiar with its regional speech.
A useful test goes beyond one polished greeting. Ask the assistant to pronounce a local name, explain an everyday concept, switch briefly into English and then return to the original language. Listen for accent drift, misplaced stress and changes in speaking rhythm. DIY AI’s guide to Spanish accent text-to-speech tools explains why selecting a voice labelled “Spanish” is not the same as choosing one appropriate for a Mexican audience.
New voices also don’t guarantee users can request any arbitrary accent, or that every dialect is supported equally well. OpenAI’s announcement does not establish either capability.
What the update means for ChatGPT, Codex and dots
| Product | Practical way to judge the improvement |
|---|---|
| ChatGPT Voice | Can a learner correct pronunciation, pause and switch languages without distracting announcements? |
| Codex | Can someone explain a coding change aloud, interrupt with a correction and receive an accurate account of what was done? |
| Dots | Does a spoken progress update make it clear what work has actually been completed? |
Codex and dots make one limitation especially important: polished speech is not evidence that an underlying task has succeeded. A voice assistant may sound assured while the actual operation is still running or has failed. Check the result, rather than treating a fluent spoken confirmation as proof of execution.
GPT-Live-1.1 is different from the original GPT-Live launch
OpenAI introduced GPT-Live in July 2026 with full-duplex conversation, meaning it could listen while speaking, respond to interruptions and delegate more demanding questions. Those are existing GPT-Live capabilities, not features first introduced by 1.1. The October release focuses on voice choice, accent quality and the frequency of spoken status updates.
The API distinction is worth clarifying. OpenAI released GPT-Live-1 through the API in September 2026. OpenAI’s new “coming soon” statement concerns GPT-Live-1.1, not the first developer availability of GPT-Live. No version 1.1 API launch date or pricing has been confirmed in this announcement.
How to test the update without relying on a polished demo
- Keep the request consistent. Use a short bilingual conversation with a regional place name and an English product name. Do not change the wording between attempts.
- Switch languages. Check that the intended accent returns after the model speaks English, instead of carrying over an inappropriate accent.
- Ask something requiring a genuine lookup. Notice whether the assistant repeats “checking”, whether it leaves an unexplained pause and, most importantly, whether the final answer is correct.
- Interrupt midway. Add a correction and check that the next answer reflects it without restarting unnecessarily.
Repeat these checks across several conversations before drawing conclusions. For recorded narration, a different assessment is needed: DIY AI’s AI audio tools comparison covers voice realism, editing control and production workflows. Live dialogue and exportable text-to-speech are different tasks, even when both sound convincing.
When will users get GPT-Live-1.1?
OpenAI says the rollout starts today, 15 October 2026, for all users of the supported Voice experiences. That doesn’t guarantee every account or app will receive the change at the same time. The announcement does not identify a dedicated version selector, so a single improved conversation is not proof of the precise model running behind it.
The next useful evidence will be consistent regional pronunciation across longer exchanges and fewer unnecessary status announcements, without losing clarity about actual task progress. Both matter more than a short voice sample.
