Voice dictation

Voice dictation turns speech into text in the message box. Press the mic, say what you want, press stop, and about a second later your words appear where your cursor is. You read them over, fix anything Perch misheard, and press Send yourself. Dictation never sends a message on its own.

Where it works

• Perch AI Web at chat.perchai.app, including on your phone and in the Home Screen app.
• Perch Terminal on the web at chat.perchai.app/agents/desktop.
• Perch Desktop on macOS and Windows. On a Mac you need version 2.4.2 or later, which is the first one that can ask macOS for the microphone. Older versions show Update the desktop app to use dictation.

Dictation is not in the CLI.

Dictate a message

To throw a recording away instead, press the X beside the stop button, or press Esc. If you are typing Chinese or Japanese with an input method, Esc closes the input method's suggestions as usual and leaves the recording alone.

Where the text goes

The transcript goes in at your cursor, so you can dictate into the middle of a message you have already started. If you have text selected, the transcript replaces it. If the message box is empty, the transcript becomes the whole message.

Perch adds one space between your existing text and the new words when they need it. Next to Chinese, Japanese, Korean, or Thai text, it adds none.

While you record

• One recording can run for 2 minutes. The timer changes color at 1:45. At 2:00 Perch stops and transcribes what you said. For a longer message, dictate it in parts.
• You can dictate while a reply is still coming in. Your next message waits in the box until you send it.
• Leaving the conversation cancels the recording. That includes switching to another thread or closing the page.
• If your microphone disconnects mid-sentence, for example a headset unplugged, Perch transcribes what it heard up to that point.
• On a phone, Perch keeps the screen awake while you record, where the browser allows it.

Languages

Dictation understands 63 languages and works out which one you are speaking on its own. You do not have to set anything up.

Afrikaans, Arabic, Armenian, Assamese, Azerbaijani, Bengali, Bosnian, Bulgarian, Cantonese, Catalan, Chinese, Croatian, Czech, Danish, Dutch, English, Estonian, Filipino, Finnish, French, Galician, German, Greek, Gujarati, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Italian, Japanese, Kannada, Kazakh, Korean, Latvian, Lithuanian, Macedonian, Malay, Malayalam, Marathi, Nepali, Norwegian Bokmål, Odia, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Serbian, Slovak, Slovenian, Spanish, Swahili, Swedish, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Vietnamese, and Welsh.

Choose your language

Some languages sound alike enough that detection picks the wrong one. Spoken Urdu, for example, can come back as Hindi in Devanagari script. If dictation keeps writing the wrong language or script, tell it which one you speak:

• Perch AI Web: open Settings and choose a language under Dictation.
• Perch Desktop: go to Settings → General → Dictation.

The choice is saved to your account, so it follows you between the web and Desktop. Choose Detect automatically to go back to detection. When a dictation comes back in a different language from the one Perch itself is set to, a short hint under the message box points you here.

Chinese characters

When Perch itself is set to Traditional Chinese (繁體中文), Chinese speech is written in Traditional characters. With Perch set to any other language, Chinese speech comes back in Simplified characters.

Names and technical words

Perch tells the speech service to expect a fixed list of words that are easy to mishear: Perch's own names, model names such as Claude, Gemini, and DeepSeek, and developer terms such as GitHub, TypeScript, and npm. The list is the same for everyone. Nothing from your chats or files goes into it.

What it costs

Nothing during the beta. Dictation does not use your Perch Tokens and does not count toward your plan's usage limits. It works the same when your chat runs on your own key, and it does not touch your key.

There are fair-use limits per account:

If you dictate many times in a row, Perch says Too many dictations right now; wait a moment and try again. The daily limit counts the last 24 hours, so room frees up as earlier dictations age out.

Privacy

• The microphone is on only while you record. Perch opens it when you press the mic and closes it when you press stop or cancel.
• Your audio is never saved. The recording lives in memory in your browser or the desktop app. It is sent once to be transcribed and is never written to disk, stored in a database, logged, or used for the Starter Data Program.
• Perch does not store the transcript. The text goes straight back to your message box and nowhere else.
• What Perch records is that a dictation happened, how long it was, and which language was detected, so it can measure the beta. Never what you said.

Once you press Send, the message is handled like anything you type.

When dictation does not work

The message under the message box tells you what went wrong. When Retry appears beside it, Perch kept your last recording and sends it again, so you do not have to repeat yourself. That copy is dropped as soon as you record again or leave the page.

If the words come out in the wrong language or script, see Choose your language.

What it does not do yet

• Text does not appear while you speak. It arrives about a second after you press stop.
• There is no hold-to-talk or keyboard shortcut. Press the mic to start and press stop to finish.
• Perch does not talk back. Dictation is for writing messages. Replies stay as text.
• It is not in the CLI.

Perch on your phone→

Add Perch to your Home Screen and dictate from there.

Perch Desktop→

Download the latest version for macOS or Windows.