Dictation and translation

In the “General” tab you choose the speech model as a tile: “Flow Air” (lightning fast) and “Flow Pro” (more precise) run LOCALLY and cost nothing; how much memory each one needs is shown on the tile. Flow Air understands 25 European languages — NOT Japanese, Korean or Chinese; if you dictate in those, use Flow Pro or Flow Max. “Flow Max” runs in the cloud, needs internet and costs 1 credit per started 3 minutes of recording, but in return it's the most accurate and the fastest to come back. On weaker devices, Flow Air is the better choice. A selected local tile stays loaded, so the first sentence arrives right away; there's no longer a switch for that. Clicking the selected tile a second time deselects it and frees up the memory — without a speech model, the dictation key is idle.

chatFlow is meant to run on older, less powerful Macs too: small before big. You choose the tier yourself. When you switch to a local model, the app first checks the free memory — if it isn't safely enough, it warns you, and you decide whether to switch anyway. If things stutter or take too long, pick a smaller tier or the cloud.

Accuracy over speed: in the “General” tab, choose the “Flow Pro” tile under the speech model (more precise, a bit slower than Flow Air).

In the “General” tab, under “Recognition Language”, you set a fixed language or “Auto-detect”. The list shows exactly the languages of the selected tier. A fixed language takes effect with Flow Pro and Flow Max; Flow Air always detects the language on its own and can't handle Japanese, Korean or Chinese.

In the “General” tab, the “Input volume” row shows how loud your microphone is set. “Check level” measures for five seconds while you speak normally. Afterwards it shows “Good”, “Too quiet” or “Too loud”. With “Too quiet” or “Too loud”, chatFlow suggests a better value: “Set to X %” applies it, “Open settings” takes you to System Settings, “Check again” measures once more. After “Good”, the row goes back to normal by itself after five seconds. The same test is available in the introduction under “Microphone”, as soon as microphone access has been granted.

While you dictate, chatFlow mutes your Mac's playback — music, videos and chatFlow's own sounds too. Afterwards everything carries on at exactly the same volume; the volume itself isn't changed. If your sound was already muted before, it stays muted. If music or a video is playing at the time (Spotify, Apple Music, YouTube in the browser), chatFlow also pauses it and lets it continue from the same spot after dictation. Only what macOS currently lists as playing gets paused; a second source is just muted. Turn it on and off in the “General” tab, row “Mute music while dictating”.

The finished text is always pasted at the cursor. If the cursor isn't in a text field (say, in Finder or on the desktop), chatFlow pastes nothing: the text then sits in the buffer, “No text field — the text is in the buffer.” appears at the bottom, and you get it out at the spot you want with the buffer hotkey (default: ⌘ Right). The buffer keeps the last text until you quit the app. Optionally, Enter is pressed after pasting (“Send automatically”, handy for chat fields), with a delay if you like. This can be set separately for individual apps.

With the cancel word you discard the sentence so far while you're speaking: everything before the last cancel word is dropped, only the part after it stays. You set the word (or several, comma-separated) in the “Styling” tab in the “Cancel command:” row; clearing the field turns it off.

Without internet, the local tiers keep working — recognition, text rules and local translation, with no credits. Everything that goes through the chatFlow cloud needs internet: the cloud tiers, the AI voice, the image command, audio processing in the studio, the speech analysis, voice training, Spark, polishing, the comma check for saved texts, Agency and chat, and the assistant Flo. Your subscription status is also checked online.

With the translator, your dictation comes out directly in the target language. You set source and target language in the “General” tab in the translator section; there's a dedicated translator key, so you don't have to keep switching back and forth. There are four tiles, and they cost different things: “Flow Air” (light, local) and “Flow Plus” (best quality, local) run on your device and cost nothing. “Flow Pro” runs in the cloud; the translation itself has no separate price, because it rides along in the same cloud recognition that transcribes your dictation — you pay for the recording at 1 credit per started 3 minutes. Important: if your speech model is already set to Flow Max, Flow Pro costs nothing extra. If it's set to one of the local models instead, it's choosing Flow Pro that creates the cloud recognition — and then the recording, which was free before, costs credits. “Flow Max” is a separate cloud call and costs 3 credits per started 500 characters, per translation with no carry-over. If your speech model is also set to Flow Max, you get 1 credit of the recording back: the translation covers it. Both cloud tiles need internet. Flow Max is available on the Pro and Agency plans; on the Solo plan (and during the trial) the tile says “Pro plan only”. The price is shown on every tile.

If you set both your speech model AND your translator to Flow Max, you used to pay twice for a single dictation: once for recording, once for translating. Since 7 August 2026, the translation covers one billing unit of the recording — 1 credit is credited back. Your account ledger then shows a credit of 1 with a note that the translation covers the recording. For a short sentence, you pay 3 instead of 4 credits in total. Exactly ONE billing unit is waived, not the whole recording. So if you record for ten minutes straight and only have one sentence translated, you get 1 credit back and pay for the rest. Without this limit, a stuck record key could make a long recording almost free. The waiver only applies if the recording cost anything in the first place — with a local speech model there's nothing to waive, because recording is free there anyway.