Dictation in ChatGPT and Claude: Speaking Your Prompts

ChatGPT and Claude both have a microphone button that writes your speech into the message box as text. You read the words, fix them and then send. That is different from the spoken conversation modes in both apps, where the app talks back. The built-in buttons work only inside that app. If you want speech in every program, a system-wide tool is the better fit, and we cover that at the end. Yap, which we make, is one such tool.

ChatGPT and Claude both turn speech into an editable prompt

Both apps keep the words editable until you send them. In ChatGPT, the control is a microphone icon in the text box. OpenAI's Academy resource gives the steps: "Tap the microphone icon in the text box and start speaking. Your words will appear as text." ¹ Claude's dictation is available on the web, on mobile and in the desktop app. ²

Editing is the reason the button is worth using, and it is also the step that is easy to skip. Speech can be quicker, but in one study it was not cleaner. In a controlled phone study by Ruan and colleagues, speech left slightly more errors in the final text than the keyboard did: 1.30% uncorrected errors against 0.79% ³. A dictated prompt needs a read-through before it goes out, and the edit-before-send step is where you do it.

The table below shows where each app's control lives, using the sources above. Where we could not confirm a surface from a source we read, the cell says so.

Dictation and voice mode are different features

Use dictation when you want to read and correct the words before the model answers. Use voice mode when you want a spoken exchange with replies you hear.

OpenAI draws the line in its ChatGPT Voice help article: "Use ChatGPT Dictation when you want to record a prompt, review and edit its transcription, and then send it as text." ⁷ Anthropic's help center draws the same line for Claude: "Dictation converts your speech to text so you can type prompts by speaking." Its voice mode is a two-way spoken conversation. ⁸

The two features are easy to confuse because both use a microphone. Dictation puts text in the box to review. Voice mode is the conversation, and it is a separate control.

If you are not sure which one you need, ask whether you want to read the prompt before the model sees it. If yes, use dictation.

ChatGPT's microphone button is documented for the phone and desktop apps

The steps in OpenAI's resource are for the ChatGPT mobile and desktop apps. Here is the sequence:

1. Open the ChatGPT app on your phone or desktop, and start a new chat or open an existing one.
2. Tap the microphone icon in the text box. ¹
3. Speak. Your words appear in the box as text.
4. Read the text, correct any mistakes, and then send the prompt.

The fourth step matters most, because the transcript is a draft of what you said, not a finished prompt.

The web version is not confirmed by OpenAI's pages

OpenAI's pages that we could read name the mobile and desktop apps for dictation, and they do not confirm the same button on the web. Test it in your own browser before you depend on it. If the button is missing or silent, check the site microphone permission first, because a browser that has been blocked from the microphone shows no activity.

ChatGPT dictation changed in January, June and August 2026

OpenAI's release notes include three dictation changes this year. On January 12, 2026, OpenAI said it was "significantly reducing empty transcriptions and improving accuracy" for all logged-in users. ⁹ On June 26, 2026, it began rolling out "a new speech-to-text model for dictation in ChatGPT across all plans", and said there are no changes to how you use dictation. ¹⁰ On August 7, 2026, it added an iOS retry for dictation that fails on a connection problem. ¹¹

OpenAI also gave a number for the June model. Its own evaluations found that word error rate was at least 10% lower for the top languages it tested than with the previous production model. ¹² That is a vendor's internal test. The note does not name the languages, the test set or the previous model, so outsiders cannot repeat it.

Which speech model ChatGPT uses

The release notes do not name the model, so we cannot say which one ChatGPT dictation uses. What we can check is OpenAI's developer guide, which names whisper-1 for one job: "To translate a completed audio recording into English, use /v1/audio/translations with whisper-1" ¹³. That is an API feature for developers. It is not a description of the ChatGPT button, so it does not answer the question for ChatGPT. Our question on which speech model ChatGPT uses lists everything we checked.

Why ChatGPT dictation fails

The reports people file fall into three groups. The first is nothing happens: the microphone responds but no text appears. A 2025 user in the OpenAI community said that "On Chrome, nothing seems to register at all." The same user said that in the app, "it listens, registers voice, but then never delivers anything to text." ¹⁴ That is one user's report from April 2025, so it shows a failure pattern rather than how common it is.

The second is a dropped connection. On iOS, ChatGPT "automatically retries once" if dictation fails "because of a connection issue or timeout." ¹¹ The retry suggests that iOS dictation needs a working connection, which matters when a long prompt is interrupted.

The third is an empty result. OpenAI's January 2026 change was aimed at this. ⁹ If you still see it, check which microphone your computer is using, and try a short test prompt before a long one.

For step-by-step fixes, see our question on why ChatGPT dictation is not working, and the general guide to dictation that is not working.

Claude's dictation works in its apps and in Claude Code

Claude has dictation in its mobile apps, in the desktop app and in Claude Code. Each has its own steps, and the phone instructions are where the sources disagree.

On phones, the Claude help article and the tutorial disagree

Claude's mobile help article gives the steps for iOS and Android:

1. Open Claude for iOS or Android and start a new chat.
2. Tap the microphone icon on the right side of the chat input field, and speak. The first time, choose your language. ^{4 15}
3. Tap the arrow icon to send the prompt. ⁴

The Claude Academy tutorial describes phone use differently: "On your phone, you hold the mic while you talk and let go to send." ¹⁶ The help article is the more formal source, so check which behavior your app shows before you rely on either.

The mobile apps dictate in a set of languages. The help article lists 12 of them ¹⁵ and marks support for languages other than English as beta. ¹⁷ Claude Code supports a longer list of 20 languages, from Czech to Ukrainian. ¹⁸ The two lists differ, so check the one for the product you use. For the full picture, read our question on Claude dictation languages.

Claude's API takes text and images, not audio

Developers looking for a speech-to-text service will not find one in Claude's model list. The models overview says that "All current models support text and image input, text output, multilingual capabilities, vision, and tool use." ¹⁹ Audio is not on that list, so transcription has to come from a separate speech service before the text reaches Claude.

Mac and desktop: a Caps Lock shortcut

On a Mac, the desktop app's quick entry can start dictation with the Caps Lock key. ⁵ The voice shortcut is optional, and you must turn it on in the app first. Voice dictation in quick entry needs macOS 14 or later. ²⁰

Claude Code: turn it on with /voice

In Claude Code, dictation is turned on by running /voice in the command line. ⁶ For the terminal permission steps and the rest of the setup, see our guide to voice input for Claude Code and AI coding tools.

Failure reports for Claude dictation

Claude dictation has public failure reports too. A user opened a bug report on June 4, 2026, which says that in the Claude Desktop app, the microphone button "never produces any transcription on macOS." ²¹ It was closed as not planned on July 10, 2026. ²¹

A second report, from a Windows user, describes dictation that works only once: "The first dictation attempt after launching the app works correctly. Any dictation attempt after that, in the same session, fails to activate or produce a transcript." ²² That report was also closed as not planned, on September 2, 2026. ²² Both are single user reports, so they show what can break, not how often. Our question on why Claude dictation is not working lists the steps to try first.

Claude's audio is processed off your device; OpenAI does not say where ChatGPT's is

For Claude, the clearest statement comes from Claude Code's documentation, which says the dictation audio "is not processed locally." ²³ So the transcription happens somewhere other than your computer. The mobile help article makes a different promise about the recording afterward: "After converting your speech input to text, we will delete your audio recording. We will not retain a copy of your speech input or use it for training our generative models." ²⁴ These are different products with different terms. The mobile article says the recording is deleted once it is converted, and it does not say where the conversion happens.

For ChatGPT we have less. The iOS retry ¹¹ suggests that dictation depends on a connection, which is consistent with transcription on a server. The release notes do not say where the audio is processed or how long it is kept, and OpenAI's privacy terms for dictation audio were not readable for this guide, so we do not state them. If that matters to you, read OpenAI's data controls before you dictate anything sensitive. Our guide to whether voice to text is safe covers how to think about this for any dictation app.

In one controlled study, speech was faster than phone typing

In the Ruan study, English speech reached 153 words per minute against 52 on the phone keyboard, which is 2.93 times faster. ^{25 26} Mandarin showed the same pattern: 123 words per minute by speech against 43 on the Pinyin keyboard. ^{27 28}

The gap changes what a prompt costs. At 153 words per minute, 52 words take about 20 seconds to say. ²⁹ A 2005 dictation study by Feng, Karat and Sears puts people's normal speaking speed at 125 to 150 words per minute, so dictating at the Ruan rate is close to ordinary talk. ³⁰ A prompt that would take a minute to type can be spoken in much less time, which makes it easier to include the background, the constraints and an example.

We did not find a study that compares the quality of spoken and typed prompts, so the case for longer prompts is practical, not measured. The Ruan numbers also come from short phrases read aloud in a quiet laboratory room, not from long AI prompts. For more on how speech fits into AI work, read our guide to speech to text for AI and the research on speaking and typing speed.

A system-wide dictation tool works in every AI app at once

The built-in buttons in ChatGPT and Claude work only inside those apps. A system-wide dictation tool types into whichever app has the cursor, so one method covers the chat window, an email and a document. Our guides on dictating into any app and on dictation in Word cover that approach in more detail.

A system tool is worth the setup when your prompts run long. On a Mac, Apple's built-in Dictation has no length timeout. Apple's current guide to Mac dictation says, "You can dictate text of any length without a timeout." ³¹ Built-in tools are enough for many people, so start there and add a separate tool only when you need one.

Yap is the other option we would point you to for system-wide dictation. It is an on-device dictation app for Mac and Windows. You can compare it with the built-in buttons on the download page, and our guide to choosing a dictation app sets out the options.

For the related questions, see why ChatGPT dictation is not working, why Claude dictation is not working, the difference between ChatGPT dictation and voice mode, which speech to text ChatGPT uses and Claude dictation languages. The apps and browsers overview links every other guide in this group.