How to Dictate to ChatGPT (2026): Every Method, and Why It Usually Is Not Broken

Yes, you can dictate to ChatGPT. Put your cursor in the message box, press the microphone icon, talk, and your speech comes back as text sitting in the box, where you can edit it before you send. Everything below was checked against OpenAI's own help pages, product documentation and release notes on 27 August 2026, with the source named each time.

By the VoiceBoard editorial teamUpdated August 28, 2026

The three voice controls in ChatGPT

ChatGPT has three things that listen to you. They have separate buttons, separate platform support and separate rules about what happens to your recording. Almost every guide on this topic treats the first two as one feature and never mentions the third.

OpenAI states the fork itself, in a section of its Voice help page headed "When should I use Voice instead of Dictation?":

Use Voice for a live, back-and-forth conversation or to talk through ideas. Use ChatGPT Dictation when you want to record a prompt, review and edit its transcription, and then send it as text.

The dictation page is blunter about the mechanics: "When you press the microphone icon to record an audio message, the recorded audio is sent to our models to be transcribed. The transcription is returned as text, which you can edit before sending as a user message."

How to tell the buttons apart without guessing. Dictation lives inside the message box, next to the attachment and tools buttons, and it leaves you looking at the same chat screen. Voice takes over the screen. Record only exists in the macOS app, only on Plus, Pro, Business, Enterprise and Edu, and it caps a session at four hours. If you are on a phone and something offered to record for four hours, you are not in dictation.

Where the dictation button is on your device

iPhone and Android

Tap the microphone in the message box, speak, and the transcription lands in the box. It does not send by itself unless you turned that on. OpenAI shipped an auto-submit setting for transcriptions on iOS on 28 February 2025, and the release note says it is off by default. So if your dictated messages keep flying off before you can read them, that switch is the first place to look.

Two Android specifics worth knowing, both from OpenAI's release notes. Since 27 February 2026 the dictation view keeps your text visible and updates it as you speak instead of showing everything at the end. Since 14 August 2026 the dictation button stays available after you have already typed something or attached a file, so you can add to a message you started. That second change closed a complaint that had been open on OpenAI's own forum since July 2025, where a user described the button disappearing "as soon as any text is entered into the input box (though it remains available in the browser version)". There is no equivalent note for iOS.

The desktop app on macOS, Windows and Linux

This is the part nobody publishes, and it is in OpenAI's own keyboard shortcut reference.

The prompting documentation spells out the behaviour: "In the ChatGPT desktop app, press Ctrl+Shift+D while the composer is visible, then start talking. ChatGPT transcribes your speech into the composer so you can review and edit it before sending the prompt."

One correction, because this gets repeated wrongly. Option + Space on macOS and Alt + Space on Windows are not dictation. They open the Chat Bar, the small launcher window. If you have been pressing those and wondering why nothing transcribes, that is why.

The Voice documentation adds a rule that explains a common confusion on desktop: "A chat or task must begin in voice mode to use ChatGPT Voice. Chats or tasks that start in another mode offer voice dictation instead." So in an ordinary chat the desktop app gives you dictation, and voice chat has to be started deliberately.

Note also that there are now two desktop apps. The rebuilt one, combining Chat, Work and Codex, arrived on 9 July 2026, and the previous app was renamed ChatGPT Classic. Linux joined on 14 August 2026 in public preview. The shortcuts above are documented for the current app.

The browser

Here is an honest gap rather than an answer. OpenAI does not document a dictation button on chatgpt.com in a desktop browser. We looked for one specifically: every dated mention of dictation in the release notes is tagged iOS, Android or desktop app, and the web instructions describe only the Voice icon in the prompt window. That is not the same as saying the button does not exist for you, since features roll out unevenly, but it does mean nobody can promise you it will be there.

If you are dictating into ChatGPT from a laptop browser, the reliable path is your operating system's own dictation, which is covered next.

What happened to Voice on the Mac app

On 11 December 2025 OpenAI announced it was retiring the Voice experience in the ChatGPT macOS app effective 15 January 2026, and said Voice would continue on chatgpt.com, iOS, Android and the Windows app. Voice returned to desktop on 23 July 2026, inside Work and Codex. Dictation was never part of that removal, which matters because "Mac app dictation no longer working" is a live search and the answer people find is usually about Voice.

What changed in ChatGPT dictation recently

Dictation has been quietly rebuilt over the past year. If you tried it in 2025 and gave up, you tried a different thing.

The June update is the significant one. OpenAI says word error rate "was at least 10% lower for top languages tested than with the previous production model", and lists gains for Japanese, Korean, Chinese, Urdu, Vietnamese, accented English and long recordings in Spanish, French, Italian and Portuguese. It also claims better behaviour in noisy public places, with quiet or whispered speech, and on strings that mix letters and numbers.

Two consequences for anything else you read on this topic. First, articles that tell you ChatGPT dictation "runs on Whisper" are describing the situation before 26 June 2026; OpenAI called the replacement "a new speech-to-text model" and did not name it. Second, the two release notes about empty transcriptions and about connection failures are OpenAI confirming, in writing, that the failures people complain about are real.

What OpenAI does not publish

Worth knowing before you trust a specification you read somewhere.

• The name of the current recognition model. Not disclosed since June 2026.
• A language list for dictation. There is none. There is a setting, Settings then Voice then Language, and a list of languages where the June update claimed improvements. That is all.
• A maximum recording length. Not documented anywhere. If you see "four hours" quoted, that is the limit for Record, a different feature.
• Automatic punctuation. Not documented as a feature or as a setting, despite being described as one all over the web.

We would rather tell you a fact is missing than fill the gap with a guess.

One thing the documentation contradicts itself on

If what happens to your voice matters to you, this is worth two minutes. Two live OpenAI help pages say opposite things.

The Voice Dictation FAQ, updated around 30 July 2026, says: "Audio from dictation will be retained for as long as the chat is part of your chat history. When you delete the chat, we'll also delete the associated audio clip within 30 days", and "If you are a consumer user and have chosen to share audio to improve our models, we may train on audio from dictation."

The ChatGPT iOS App FAQ, which showed as updated two days before we checked, says the opposite twice: "No. We send audio clips from our speech-to-text feature to our servers so that we can use our Whisper API to transcribe them, but we don't retain voice clips beyond what's necessary to complete the transcription", and "No. We only use the audio clips to prepare a transcription… The clips are then deleted."

We are not going to tell you which page is right, because OpenAI has not. The dictation page is the one written specifically about dictation; the iOS page is the one still describing Whisper, which the June model change makes questionable. The practical move, if this matters to you, is to open Settings, then Data Controls, and turn off "Include your audio recordings", or turn off "Improve the model for everyone" entirely. That toggle covers dictation and voice chat together. Business, Enterprise and Edu workspaces are excluded from training by default.

Dictating with your own device instead

Your phone and your computer already have dictation, it types into any text field including ChatGPT's, and on several platforms it never sends your voice anywhere. That last part is a real advantage over ChatGPT's own microphone, which is a cloud service and stops working the moment your connection does.

iPhone. Settings, General, Keyboard, Enable Dictation. Then tap the microphone on the keyboard whenever a text field is open. Apple's documentation is unusually direct: "Dictation requests are processed on your device in many languages, no internet connection is required." Punctuation is inserted automatically, with a toggle called Auto-Punctuation if you want it off. There is no length limit, but dictation stops on its own after 30 seconds of silence, and the keyboard stays open the whole time so you can fix a word by hand and carry on speaking.

Mac. System Settings, Keyboard, Dictation. Start it with the microphone key, a shortcut you choose, or Edit then Start Dictation. Press and release the microphone key for dictation; hold it and you get Siri instead, which trips people up. Apple states you can dictate text of any length without a timeout, with the same 30 second silence cutoff (Apple Support). On an Apple silicon Mac, supported languages are processed on the machine. On an Intel Mac, or in a language without on-device support, your speech goes to Apple's servers. If you have Voice Control switched on, standard Dictation is unavailable.

Android. Tap the microphone at the top of Gboard, wait for "Speak now", and dictate punctuation by saying "period", "comma", "new line". Google's page carries a caveat worth repeating: "Talk-to-text doesn't work with all languages", and no list is published. The better version, advanced voice typing, needs a Pixel 6 or later, shows a colourful microphone icon, and comes with a strong privacy claim: "The text you speak stays on your device and isn't sent to Google servers", with exceptions for the Fix it and detailed edit features.

Windows. Press Win+H with your cursor in the box. Microsoft is clear that this is not local: "Voice typing uses online speech recognition, which is powered by Azure Speech services", and you need to be connected. The published language list held 46 variants when we counted it on 27 August 2026, including Russian, and automatic punctuation is a toggle in the panel's settings. For offline work there is Voice Access, on Windows 11 22H2 and later, which recognises speech on the device and runs "without an internet connection" after a one-off language download. It covers 15 locales only, and its automatic punctuation is off by default. Our full comparison of Windows options is here: best dictation software for Windows, and the Mac equivalent is best dictation software for Mac.

The short version: on an iPhone, an Apple silicon Mac or a Pixel, your operating system will dictate into ChatGPT without your voice leaving the device. ChatGPT's own microphone will not. That is the answer to dictating on a plane, and it is also why dictation goes silent on a weak connection while the rest of the app still works.

When the built-in microphone runs out of road

Built-in dictation is enough for most people most of the time, and if it is enough for you, there is nothing else to do. Three situations are where it stops being enough.

The first is scope. ChatGPT's microphone dictates into ChatGPT. If you also want to speak into Slack, your email, a terminal or a code editor, you are learning a different control in every app. The second is cleanup. Built-in dictation transcribes what you said, including the false start and the "um" and the sentence you abandoned halfway. The third is vocabulary. Names, product names and industry jargon come out wrong every single time, forever, because none of the built-in tools learn yours.

Tools that sit on top of any text field solve all three at once. The field is bigger than the marketing suggests, and it includes free options.

• Voice In is a Chrome extension with a free tier, which makes it the lowest effort thing to try if you live in a browser.
• Spokenly advertises free use with local models and no account, on Mac.
• Voicy covers Windows, Mac and Linux plus a Chrome extension, from $8.49 a month with a $260 lifetime licence, and is one of very few options with a real Linux build. It runs through the cloud, with no offline mode.
• VoiceInk and Superwhisper run models locally on a Mac, which is the only way to get dictation into ChatGPT with your audio never leaving the machine (the Mac roundup covers both).
• Wispr Flow is the one we point most people at, because it does the cleanup part best and because one account follows you across Windows, Mac, iPhone and Android. You hold a key, you talk the way you actually talk, and finished text appears in whatever app you were in, filler removed and punctuation in place. It picks up names and jargon as you use them, so the second week is better than the first, and it covers 100+ languages. Our full write-up, including what we do not like about it, is the Wispr Flow review.

The fine print on that last one, since we are recommending it. It is a cloud service, so it fails offline exactly like ChatGPT's own microphone does. On Windows it needs 10 or 11 on x64; the documentation states that Windows on ARM devices, meaning the Snapdragon laptops, are not supported. The free tier is 2,000 words a week on desktop and 1,000 a week on iPhone. Pro is $15 a month, or $12 a month billed yearly.

Try Wispr Flow

ChatGPT dictation not working

Search this phrase and you land in forum threads, some of which have been open since 2025 with no reply from OpenAI. Here is what is actually going wrong, ordered by how often it turns out to be the answer, with a dated source for each so you can check us.

1. You pressed the wrong control

If ChatGPT answered you out loud, or the screen changed and there was no text in the box, you were in Voice, not dictation. This is not a small edge case. On OpenAI's own forum, participants in a thread running from April to September 2025 reported precisely the split: the voice conversation worked, dictation did not. Before you change a single setting, look for the microphone icon inside the message box and press that one.

The same confusion runs backwards. People search for dictation, find articles about Voice, conclude the feature was removed, and go looking for a replacement. In March 2025 a user opened a thread titled "Speech-to-text (microphone) button gone. Only Voice chat available", which is the same collision from the other side.

2. Microphone permission, at more than one layer

There are usually two or three gates, and fixing one does not help if another is closed.

• iPhone: Settings, ChatGPT, Microphone. Then check Screen Time, Content and Privacy Restrictions, Microphone, which can stop apps requesting access at all.
• Android: Settings, Apps, ChatGPT, Permissions, Microphone.
• Windows: Settings, Privacy and security, Microphone. Three switches matter: microphone access overall, access for apps, and the app itself.
• Mac: System Settings, Privacy and Security, Microphone.
• Browser: the site permission is separate from the system one. In Chrome, Settings, Privacy and security, Site settings, Microphone, and set chatgpt.com to Allow rather than leaving it on the default. A long-running OpenAI forum thread, open from November 2024 into February 2026, is full of people who had granted system access and still saw "Enable microphone access in Settings" until they set the site permission explicitly.

3. Your browser, specifically

If dictation fails in one browser and works in another on the same machine, it is the browser.

The clearest documented case is Brave. On 6 July 2026 a user reported that clicking the microphone in the ChatGPT message box immediately showed "Content failed to load", with InvalidStateError: Failed to execute 'showPopover' on 'HTMLElement' in the console, while the same account worked in Chrome on the same computer. A Brave staff member replied the same day linking the bug tracker. The issue, brave/brave-browser #56858, was opened on 3 July 2026, has confirmations from macOS and from Linux Mint, is labelled for both desktop and Android, and was still open when we checked it on 27 August 2026.

The original reporter added a line that belongs in this guide: "This is not the full Voice Conversation (blue orb) interface." The bug hits dictation and leaves Voice alone, which is exactly the symptom that sends people down the wrong diagnostic path.

If you are on Brave, the fix today is to dictate in Chrome or Edge, or use your operating system's dictation instead.

4. It is OpenAI, not you

Check status.openai.com before you rebuild your microphone setup. Dictation has had its own named incident: "ChatGPT Dictation Elevated Error Rate", a partial outage on 2 April 2026 running from 14:17 to 17:49. A second one, "Elevated transcription failures affecting ChatGPT & Codex", ran from the evening of 7 May 2026 to the morning of 8 May. No article in the search results for this problem mentions the status page.

5. The connection dropped, or it timed out

Dictation is a cloud service by design, so a weak connection produces nothing rather than a warning. OpenAI acknowledged this on 7 August 2026 with an iOS change: "If dictation fails because of a connection issue or timeout, ChatGPT automatically retries once, helping you finish recording without starting over." One retry. After that you lose the recording.

6. Nothing comes back at all

An empty transcription is a known failure that OpenAI addressed by name on 12 January 2026, "significantly reducing empty transcriptions and improving accuracy" for all logged-in users. If you are still getting silence back, you are hitting the residue of a real bug, not doing something wrong.

7. Long recordings vanish

This one is unresolved and worth knowing before you dictate something you care about. A bug report opened on OpenAI's forum on 23 April 2025, still collecting replies in November 2025, describes dictation across iOS, macOS and web deleting recorded speech with no way to recover it, sometimes preceded by a timer running backwards to "-5000 seconds" and an error reading "Your dictation must be longer than 1 second". One user wrote: "I have lost all my dictations about 5 times in the last 6 hours of my use."

OpenAI does not publish a maximum recording length, so there is no threshold to stay under. The practical habit is to dictate in short passes and let the text accumulate in the box rather than recording one long take.

8. It sends before you can read it

If your dictated message fires off without showing you the text, check the auto-submit setting, which OpenAI shipped on iOS on 28 February 2025 with the note that it is off by default. Separately, an Android regression reported on 1 April 2025 sent messages immediately without displaying the transcription at all, which is a bug rather than a setting.

9. It transcribes the wrong language

Open Settings, then Voice, then Language, and pick the one you speak. OpenAI has a short help page saying exactly that. Detection is automatic otherwise, and it guesses wrong often enough for bilingual users that the June 2026 model update called out mixed-language speech specifically.

10. Someone turned it off

On a work or school account, an administrator can disable Voice for the whole workspace. On a linked teen account, a parent or guardian can manage access through parental controls. Availability also depends on plan, region and app version, so an option a colleague has may simply not be enabled for you.

11. The button needs a second click

On the Windows desktop app, a report from 9 May 2026 with several confirmations describes the microphone button not activating on the first click: "Sometimes dictation starts only after a second click." Annoying rather than serious, and worth trying before you conclude anything is broken.

Fix it on your device: the short version

How we checked this

We did not install the apps, run a stopwatch, or measure accuracy, and we are not going to imply otherwise. Read the pages currently ranking for this search and you will find confident claims about speed multipliers and language counts published by companies selling the tool that comes first in their own list, with nothing attached that you could verify.

What we did instead, on 27 August 2026, was check the things that can be checked, at the source.

• Read OpenAI's Voice Dictation FAQ, ChatGPT Voice page, ChatGPT Record page and the iOS and Android app FAQs, and quoted them rather than paraphrasing.
• Read the product documentation at learn.chatgpt.com, including the keyboard shortcut reference, which is where the Ctrl+Shift+D and Ctrl+Shift+V figures come from.
• Went through OpenAI's public release notes and pulled out every dated entry about dictation over the past eighteen months.
• Opened OpenAI's status page and recorded the two incidents affecting dictation and transcription, with their times.
• Read the Brave community thread and then queried the linked bug tracker to confirm the issue was still open on the day of writing.
• Read dated user reports on OpenAI's own community forum and quoted them with links and dates.
• Read Apple's iPhone and Mac dictation guides, Google's Gboard pages and Microsoft's voice typing and Voice Access pages for the system-level methods, and counted the Windows language list ourselves rather than repeating a number.

Three limitations, stated rather than hidden. We have no screenshots of the ChatGPT interface, and this is a topic about where to click, so the descriptions here are from documentation rather than from a screen we photographed. Reddit was not reachable from our research environment, so where community experience appears here it comes from sources you can open: OpenAI's forum, Brave's forum and a public bug tracker, all dated. And the interface in this product changes fast, with two documented redesigns of the voice experience since November 2025, so treat anything here about the position of a button as a snapshot rather than a permanent fact. We recheck this page quarterly.

What to do next

Spend a week on your device's built-in dictation and notice where it fails you. If it fails because you keep leaving ChatGPT for other apps, or because you spend longer tidying the transcript than you saved by speaking, the free 2,000 words a week on Wispr Flow is the cheapest way to find out whether a paid tool changes anything before you spend a cent. If it fails because you were offline, no paid cloud tool will help you and you want a local one instead: the alternatives worth comparing.