What’s New With GPT-Live-1.1? ChatGPT’s New Voices and Accents
beginner
The Rundown
OpenAI is rolling out GPT-Live-1.1 and GPT-Live-1.1 Quip in ChatGPT Voice. The October 15 update focuses on more natural accents, new voices, and more varied spoken cues while searching. Quip is the smaller version.
Try it on something you can check: a short work update. You will compare voices, try a regional voice, correct a detail while ChatGPT speaks, and check a spoken search answer. No app build or API key is needed for this exercise.
University Pro members get the practice prompts, listening checklist, and comparison record. The prepared files contain no voice recordings.
What changed in GPT-Live-1.1?
Change
What OpenAI announced
What to try
Accents
More natural accents across languages.
Use a short passage in a language you know well.
New voices
More language-specific choices, plus Bubbie in English.
Compare two voices using the same words.
Search cues
More variety in the short phrases Live uses while searching.
Listen during a real lookup, then check the source.
Quip
A mini version alongside GPT-Live-1.1.
Check the version available to your account before labeling a test.
These are OpenAI’s announced changes. The release includes no measured speed gain or accuracy score. Our screenshots show the available controls; the listening and spoken-interruption steps below are exercises for you to try.
Several useful features predate this update. Live already supported interruptions and search; plugins arrived on September 23. Today’s release focuses on the voice experience.
Pick a voice for comfort and clarity. You might prefer one for a short update and another for language practice. A warmer delivery does not make a wrong date correct.
Where we would use the update
These are starting ideas to test:
Task
A focused request
Practice a presentation
Ask for one question at a time about an approved outline.
Language practice
Try a short exchange and ask for one correction at a time.
Talk through a decision
State the options, then ask it to repeat your constraints before offering advice.
Get a spoken update
Ask for three findings from a named source, with links left in the chat.
Keep important names, figures, and decisions visible in writing. Choose the voice that helps you follow and check the work.
For connected tasks, our Run Your Workday by Voice With ChatGPT guide covers the wider workflow. This release guide stays focused on listening and simple checks.
Can you use the audio in a video or podcast?
Check the rights before exporting or publishing a recording. Section 8 of OpenAI’s Service Terms limits ChatGPT Voice Output to non-commercial use and prohibits distributing it as a standalone audio or sound file.
That matters for a paid training site and for anyone preparing ads or podcasts. We would not treat a ChatGPT call like a licensed voiceover export, or assume that putting the same audio in a video or behind a paywall resolves the restriction.
For a published audio demonstration, obtain suitable permission or use an official embed whose terms permit that use. Until then, follow the listening exercise inside ChatGPT. This guide’s Pro materials contain written prompts and checks. No ChatGPT Voice recordings are included.
Turn a spoken idea into work you can use
A clear voice can help you discuss a task. A good brief, sound sources, and careful review determine whether the result is useful.
Nate Grahek’s Getting Knowledge Work Done with ChatGPT Work and GPT-5.6 covers preparing source material, planning deliverables, reviewing drafts, and creating editable outputs. The recording teaches the wider knowledge-work process.
Pro members also get this guide’s written practice prompts, listening checklist, and comparison record. ChatGPT access and voice usage have separate terms.
Step 1
Open Live and choose a voice
On web or mobile, open Settings → Voice, then select Voice. If your account offers a mode selector, choose Live. In a new chat, select the Voice icon and allow microphone access, following OpenAI’s setup instructions.
In our web account, Voice settings opened directly to the voice and language controls. Bubbie and the regional voices below were present. The settings screen did not expose an exact model-version label.
Use fictional details and keep the first test separate from private work. Read your account’s usage notice before starting a long session. Rollout and workspace settings can affect access.
Choose an existing voice for the first pass. Record its name and your device. When the app exposes a model label, record that too. A model identifying itself in its reply is not evidence of which version ran.
Bubbie and the existing American English voices in ChatGPT’s real Voice settings.
Step 2
Compare the same notice in two voices
Paste or speak this request:
Prompt
Read only the passage below, exactly as written, at an easy speaking pace. Add no introduction or closing comment. This is a fictional practice notice. Use no tools or connected apps.
Our workshop starts on October twenty-second at nine in the morning. We expect twenty-four people. The total budget is nine hundred US dollars. We'll take questions at the end.
Listen for the date, time, headcount, and budget. Check whether you could follow the notice without looking at the text. Note any skipped word or odd pause.
Then choose Bubbie, when available, and repeat the same prompt in a fresh chat. This keeps the task context separate from the first read.
Keep the passage, language, device, and requested pace unchanged. This helps you compare the voices without also changing the brief.
Check
What you are listening for
Words
The requested passage, without extra lines or omissions.
Details
October 22, 9 a.m., 24 people, and $900.
Pacing
Gaps that make the notice easy to follow.
Preference
Which voice you would choose for this task.
This is a voice comparison within the current experience. It is not a GPT-Live-1 versus 1.1 benchmark. If you lack a verified older recording made under matching conditions, leave the old-model comparison out.
OpenAI’s voice evaluation guidance separates conversational quality from task correctness. Check the spoken delivery and the details in the answer.
The fictional workshop notice prepared in an unpersonalized temporary chat before a reader starts the listening exercise.
Prepare the written notice, then use a regular Voice chat for the listening exercise. This screenshot shows the unsent practice request.
Step 3
Try a regional voice in a language you know
Choose a voice for a language you can judge. Set your preferred language under Settings → Voice → Language.
ChatGPT’s language selector under Voice settings, with the personal avatar blurred.
For example, choose Tinta or Arrullo when available and ask it to read this fictional Mexican Spanish notice without adding a translation:
Prompt
Read only the following fictional notice in Mexican Spanish. Keep its facts and wording. Add no introduction, closing comment, translation, or tool use.
El taller empieza el 22 de octubre a las nueve de la mañana. Esperamos a 24 personas. El presupuesto total es de 900 dólares estadounidenses. Dejaremos las preguntas para el final.
It carries the same core facts as the English passage: October 22, 9 a.m., 24 people, and a $900 US-dollar budget. Have a fluent speaker review pronunciation and phrasing before making a claim about the accent.
Then ask:
Prompt
Repeat the notice a little more slowly, with a brief pause between sentences. Keep every fact unchanged. Do not add an explanation.
This is a conversational request. OpenAI’s GPT-Live prompting guidance recommends stating the target language and checking regional delivery with the selected voice.
Listen for preserved meaning as well as clearer delivery. A slower read that changes a number fails the exercise.
Use a language you understand for the first test. A polished-sounding foreign-language clip gives you little evidence when you cannot check its words. Have a fluent reviewer assess any other language sample before sharing a conclusion.
Tinta and Arrullo in the Mexican Spanish section of the real ChatGPT voice picker.
Step 4
Correct a detail while ChatGPT speaks
Return to the English notice. Ask for another read, then interrupt with:
Prompt
Wait, change that to October twenty-third and thirty people. The time and budget stay the same. Read the updated notice in two short sentences. This is still a draft; do not create an event or send anything.
Check the final answer against this small record:
Detail
Before
After your correction
Date
October 22
October 23
People
24
30
Time
9 a.m.
9 a.m., unchanged
Total budget
$900
$900, unchanged
Listen for whether it stops the old read and uses the corrected details. The response should remain a draft. Nothing in the exercise requires a calendar event, message, or booking.
If it misses the correction, stop and say which detail is wrong. Try again without changing several other instructions. Keep the failed result in your review notes.
For longer sessions, use a short direction such as:
Prompt
Give one point at a time. Keep each reply to two sentences unless I ask for more. When I correct a detail, repeat the corrected version before continuing. Use only the fictional notice unless I explicitly request the separate official-source lookup. Do not use connected apps, create events, send messages, or publish anything.
Use a new chat for the lookup so the sample workshop stays separate. Say:
Prompt
Search OpenAI's official ChatGPT release notes for the entry dated October 15, 2026 about GPT-Live-1.1. Tell me the date and the name of the new English voice in two short sentences. Leave the source link in the chat. If you cannot search or find that entry, say so. Do not use connected apps or take actions outside this lookup.
Listen for whether the voice signals that it is looking something up. Then inspect the actual search activity and source link.
For this specific check, the release entry identifies October 15, 2026 and Bubbie. If ChatGPT points to another entry or never performs a lookup, mark the search check incomplete. A correct answer from prior context would not establish that search ran.
A spoken “searching” cue is not proof of a successful search. Open the page and check what it says. A smooth transition into a wrong answer is still a wrong answer.
OpenAI’s voice-agent testing guidance recommends tracing the whole task, including tool activity and the final result.
Open the official release entry after your lookup and compare it with ChatGPT’s answer.
Keep going
Keep learning with Rundown Pro
Pro adds the full on-demand course library, certifications, group office hours, and eligible member perks.
$29/mo, cancel anytime
What is GPT-Live-1.1 Quip?
Quip is the smaller model announced alongside GPT-Live-1.1. Keep its results separate when comparing versions; the release provides no head-to-head quality score.
Who gets GPT-Live-1.1 or Quip?
The current ChatGPT Voice limits table lists Quip for Free and Go, and GPT-Live-1.1 for Plus and Pro. Available usage and workplace access depend on your plan. Check the current table and your account’s usage notice before a long session.
Is Bubbie another model?
Bubbie is an English voice option. Quip names the smaller model. Selecting a voice does not establish which underlying model version handled a session.
Is GPT-Live-1.1 available through the API?
The confirmed release is in ChatGPT Voice. The public API model page checked for this guide still documents gpt-live-1, and the API changelog does not establish a separate 1.1 release. We could not verify a public 1.1 model ID or price.
An API route also has its own billing and terms. Do not label an older API model’s audio as 1.1 or assume it provides the same ChatGPT voices.
What if the voice sounds good but gets a fact wrong?
Correct the fact and ask for a short readback. Check the source yourself. Keep voice quality and factual accuracy as separate judgments.
What if it responds before I finish?
Try a quiet setting and headphones, then ask it to wait until you request a response. Record whether that helps. Use the result to decide which setup works best for you.
Is a voice conversation private?
Review your account’s data controls before using work material. Consumer and business plans have different training defaults. Start with fictional content and keep people who have not agreed to participate out of the test.
