What to expect from Flow: accuracy and known limitations

Flow turns speech into text and can clean up the result. Accuracy depends on the captured audio, language, and formatting settings. Use this guide to distinguish expected behavior from a problem that needs troubleshooting.

An internet connection is required

Dictation is transcribed in the cloud. A brief connection failure can delay the result; if processing fails, retry saved audio when offered. A microphone failure is different: uncaptured audio cannot be recovered.

Reconnect before retrying. On desktop, wait for automatic recovery before dismissing the dictation or starting another—either action can stop the old result from pasting. If network or VPN restrictions interfere, see transcription error troubleshooting and network configuration.

Language and formatting can change the output

Flow supports more than 100 dictation languages, with varying accuracy and formatting quality. It detects a language for each dictation, not independently for every word. Mixed-language speech can be less reliable, particularly English mixed with Chinese or Japanese.

• Choose the intended language or variant. British or Canadian English, Chinese script choices, and Swiss German affect spelling or script. Variants in the same group are mutually exclusive.
• Choose Hinglish explicitly for Hindi-English dictation in Latin letters. Selecting Hindi and English alone does not request romanized Hindi. Some English loanwords can acquire phonetic spellings; dictionary corrections can help.
• Expect editing when cleanup is enabled. Flow can remove filler words, infer punctuation, and organize paragraphs. These changes are different from mishearing speech.

Use language settings for each platform's choices, the dictionary for recurring terms, and Auto Cleanup for desktop editing levels. Android also has language settings; it does not offer a Smart Formatting toggle.

If a long dictation's formatting cannot finish, some text can arrive unformatted. A Hinglish ending may remain partly in Devanagari. Retry when available. If the words themselves are missing, compare the recorded audio rather than assuming it is only formatting.

Long recordings can stop or take longer to process

Platform	What to expect
Mac	Up to 20 minutes, with a warning at 19 minutes and a Recover text action at the limit. See longer dictation sessions.
Windows	Follow the session warning and use Recover text when offered. Start another dictation after the previous one finishes processing.
Android	Five minutes. A warning appears about one minute before automatic submission; the limit cannot be disabled. See Android session limits.
iPhone	The Disable Flow Session setting governs background-session behavior; it is separate from recovering a saved dictation. If the keyboard is not open when a result is ready, check History.

A connection interruption near the end can add processing time or leave an incomplete result. Keep the session open while recovery runs, then use History's Retry if needed.

Text insertion depends on the app

Mac and Windows

Flow uses the clipboard to paste text, then restores supported previous clipboard content. Standard text and images can be restored; copied files, PDFs, audio, and some rich document formats are not preserved. Save important clipboard content before testing a troublesome app.

If insertion fails, follow the notification: it may offer manual paste or a Paste Last Transcript shortcut. Defaults are Cmd+Ctrl+V on Mac and Shift+Alt+Z on Windows. Use your configured shortcut if changed.

Some terminals and remote desktops need a workaround. Flow runs locally for remote sessions, and context-aware formatting is unavailable inside those sessions. See remote desktops and text not pasting.

On Mac, dictated text is concealed from many clipboard managers; Windows clipboard managers may retain it. In some editors, insertion occurs in sections, so the previous clipboard is restored only after the last section.

Android

Flow inserts into supported focused fields and can fall back to clipboard paste. If transcription succeeded but insertion failed, use Paste. If transcription failed, use Retry once online.

The bubble is excluded from sensitive fields such as passwords and payment details, and from recognized banking apps and password managers. Browser address bars and notification replies can still support dictation. Large or complex app screens can require clipboard fallback.

Recovery depends on stored audio

Dictation history is local to the recording device and does not sync across devices. Desktop audio download and retry are available only while the audio remains—normally up to 14 days, with shorter or disabled retention possible. Retry cannot run during an active dictation. Deleting a desktop history entry also removes its audio.

iPhone and Android recovery likewise require audio saved on that device. Audio capture errors cannot be retried as a transcription. Use Retry failed transcriptions for platform actions and history and deletion for retention details.

Check audio before blaming the transcript

Wait briefly for the microphone before speaking. Wireless mics can need time to wake; a distant mic, near-silence, or background noise can produce missing words or unrelated short phrases. A closed MacBook requires an external mic. Virtual or loopback inputs can capture an unintended audio source.

For an abrupt change, follow Transcription suddenly got worse. For permission or missing-input warnings, use microphone troubleshooting. Report persistent differences between clear recorded speech and the transcript, with your platform and a specific example.