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 article to tell expected behavior apart from a problem worth troubleshooting.

An internet connection is required

Dictation is transcribed in the cloud. A brief connection failure can delay the result; if processing fails, reconnect and retry the saved audio when offered. A microphone failure is different: uncaptured audio cannot be recovered. If network or VPN restrictions interfere, see network configuration.

Language and formatting change the output

Flow supports more than 100 dictation languages, with varying accuracy and formatting quality. Accuracy is strongest in widely spoken languages such as English, Cantonese, Chinese, Hindi, French, Spanish, Korean, Portuguese, and Russian. Flow records one detected language per dictation, not per word, so mixed-language speech is less reliable. Changed wording or layout in the pasted text is formatting, not misheard speech.

• Full punctuation and capitalization cover 33 languages: including Japanese, Korean, Arabic, Hebrew, Turkish, Polish, Dutch, Vietnamese, and Indonesian. Outside those languages, expect lighter or plainer formatting.

• The dominant language wins: if you switch languages mid-dictation, the language you spoke longest is recorded. History records any language that was recognized, and marks a dictation undetermined only when no language could be identified at all.

• Variants are recorded and exclusive: History shows the variant you selected — Hinglish, British or Canadian English, Traditional Chinese, Swiss German — rather than the base language, and variants in the same group are mutually exclusive. Choose the one you intend, since these affect spelling and script.

• Similar languages are double-checked: with both Hindi (or Hinglish) and Urdu selected, or both Russian and Ukrainian selected, Flow verifies the language before finishing, so Urdu speech comes back in Urdu and Ukrainian speech in Ukrainian, while genuine Hindi or Russian speech is unchanged. This adds a little processing time. With only Russian selected, nothing changes.

• Less common languages stay in the language you chose: select Czech, Hungarian, Thai, Arabic, or Greek and speak it, and Flow holds the transcript to that language instead of returning a close neighbour or the wrong script, adding roughly a second of processing time. If your selection also includes an easily confused language — Slovak alongside Czech, or Malay alongside Indonesian — Flow keeps both open so either can be recognized. Nothing changes if you have no language selected or include auto-detect.

• Cleanup keeps your language mix: mixed-language speech such as Hinglish stays in the same blend — grammar, clarity, and filler words are cleaned up without translating into English. Text is translated only if you ask for it in a custom instruction.

See language settings for each platform's choices and the dictionary for recurring terms.

Long and back-to-back dictations take longer

A long dictation can finish in several parts, each with its own processing time, so the end of a long recording is not cut short. Long dictations come back fully formatted, including ones that contain or end in bulleted or numbered lists, and Hinglish dictations. If the connection drops near the end, Flow transcribes the leftover audio on Mac and Windows and pastes it with the rest.

On desktop you can start a new dictation right after finishing one. If the earlier dictation is still being processed, the new recording waits its turn, then pastes and appears in History as usual. A dictation you dismiss with Escape while it is waiting is discarded without a result.

Text insertion depends on the app

Mac and Windows

Flow pastes text using the clipboard, so save important clipboard content before testing a troublesome app. On Mac, Flow restores what you had copied before only after the dictation lands in the app, and never overwrites something you copied in the meantime; in a busy app such as Chrome the dictation can stay on the clipboard for up to about 7 seconds, so pressing the paste shortcut during that window pastes the dictation again. On Windows, your previous clipboard content comes back about half a second after the dictation is pasted, and anything you copy during that moment is kept.

If insertion fails, paste the text yourself with Paste Last Transcript — Cmd+Ctrl+V on Mac and Shift+Alt+Z on Windows by default, or your configured shortcut. See text not pasting.

Android

Flow inserts into supported focused fields and can fall back to clipboard paste, which large or complex app screens may require. If transcription succeeded but insertion failed, use Paste; if transcription failed, use Retry once online.

Check your microphone before blaming the transcript

Wait briefly for the microphone before speaking, because wireless mics need time to wake, and confirm the microphone you intend to use is selected as the input device. On desktop, Retry cannot run during an active dictation. For an abrupt change in quality, follow Transcription suddenly got worse.

Still stuck?

Contact support if clear recorded speech consistently differs from the transcript. Include your platform and a specific example.