What to expect from Flow: accuracy and known limitations
Last updated: August 11, 2026
Available on: Mac, Windows, iOS, Android
Wondering why your first words got clipped, why a long dictation cut off, or why one language is less accurate than another? This page explains what is normal behavior, what is a hardware or platform constraint, and what is worth troubleshooting.
An internet connection is required
Flow transcribes in the cloud, so there is no offline transcription on any platform. If your connection drops mid-dictation, the transcription fails and you can retry it once you are back online.
VPNs: Flow is not compatible with most VPNs. If you need one, enable split tunneling and allow wisprflow.ai and its subdomains.
Server routing: Flow routes to the nearest server for the lowest latency and sends small background pings to find it. That is expected behavior.
Language accuracy varies by language
Flow supports 100+ languages, but accuracy and formatting quality are not identical across all of them. Language is detected per dictation session, not per word.
Fully optimized formatting: English, French, German, Hindi, Italian, Portuguese, Spanish, Thai, Japanese, and Korean.
Confident transcription: English, Spanish, Portuguese, French, Russian, German, Italian, Dutch, Japanese, Turkish, Polish, and Catalan.
Everything else: All other supported languages work, but use a general purpose formatting path rather than a language specific one. Non-English transcription continues to improve but is not yet as reliable as English.
English-only features: Personalized Flow Styles apply only when you are dictating in English, on every platform.
Tip: Select only 2 to 3 languages. Selecting more reduces accuracy, because Flow has a wider set of possibilities to choose between on every dictation.
Mixing languages in one dictation
Works well: English with Spanish, French, or German.
Less reliable: English with Chinese or Japanese. English words may be transcribed in Chinese characters, or the reverse.
Code mixing: Supported. Select Hinglish explicitly rather than relying on Hindi or English alone.
Session length limits
Each dictation session has a maximum length. When you reach it, Flow submits what it has rather than discarding it.
Platform | Limit | What happens |
Mac and Windows | 20 minutes | Warning at 19 minutes. At 20 minutes the session ends and a Recover text button re-pastes your transcript. |
iOS | 5 minutes | Auto-submits with no advance warning. |
Android | 5 minutes | Toast warning about a minute before, then auto-submits. Rolling out gradually, so it may not be active for you yet. |
Note: A "connection lost" error at the 19 to 20 minute mark on desktop is the recording cap, not a network problem.
How Flow inserts text, and what that affects
Flow pastes through the system clipboard, then restores what you had copied before.
What is restored: Standard text and image formats. File contents, PDFs, RTFD documents, audio formats, and copied files are not.
If a paste fails: Flow writes your dictated text to the clipboard so you can paste it manually, and your previous clipboard contents are not restored. Paste anything important somewhere safe first if paste has been failing in that app.
Clipboard managers: On Mac, dictated text is marked as concealed, so most third party clipboard managers will not store it in their history. On Windows it is not concealed and may appear.
Where direct paste does not work
Remote desktops such as Citrix, RDP, VMware Horizon, and Windows 365. Flow cannot be installed inside a virtual desktop and must run on your local machine. Context aware formatting also does not work inside remote sessions.
Some terminals including WSL, tmux, screen, SSH, and Termius on Windows.
In these cases, use Paste Last Transcript: Ctrl+Cmd+V on Mac or Alt+Shift+Z on Windows.
Microphone and audio expectations
Recording start: Flow discards a small amount of audio at the start of every recording, on every microphone. Pause briefly before you speak to avoid clipped first words.
Wireless microphones: AirPods and other wireless mics have a variable wake-up delay that can drop your opening words. Built-in and wired USB microphones have none.
Very short audio: Clips shorter than 0.3 seconds are skipped without an error.
Near-silent input: Flow may occasionally output a short spurious phrase instead of real speech.
Virtual microphones: Not supported. Use a physical microphone.
Mac clamshell mode: The built-in microphone does not work with the lid closed. Use an external microphone.
FAQs
The end of my long dictation looks unformatted. Did Flow drop words?
No. On dictations over about 30 seconds, the tail end can appear as plain unformatted text. Your full dictation is always pasted and words are never silently dropped.
Flow removed my filler words and changed my punctuation. Is that a bug?
That is Smart Formatting working as designed: it removes fillers, infers punctuation, and structures paragraphs. Backtrack also cleans up false starts and self-corrections, including when you restate yourself without a trigger word. On Mac and Windows, set the Auto Cleanup level in Personalization to control how much cleanup happens; None disables it entirely.
My accuracy dropped suddenly. What changed?
This is almost always an audio change rather than a model change: a different microphone, a wireless connection, lower speaking volume, or more background noise. Work through the checks in Transcription suddenly got worse or feels less accurate.
Does Flow work with a Dvorak, Colemak, or non-Latin keyboard layout?
Flow reads your keyboard layout when it starts. For non-Latin layouts such as Cyrillic, Arabic, or CJK, switch to QWERTY before launching Flow and keep it active while dictating. For Dvorak or Colemak, set your layout before launching Flow and keep it for the whole session. If you change layouts mid-session, restart Flow.
Can I recover a dictation that failed?
Usually yes. Flow saves your audio locally as you dictate, and a Retry option appears in history. On desktop, audio is kept for 14 days and retry requires at least 5 seconds of audio. Microphone errors are the exception, because no audio was captured to retry.