Speak in any language.
Paste clean text anywhere.

01 · General

System permissions

RevFly needs to hear your hotkey in any app and to paste text where you type. On macOS both depend on the Accessibility permission.

Accessibility

Click Sync. macOS opens Privacy & Security → Accessibility; switch RevFly on and the badge turns Granted. Without it, the hotkey and auto-paste won't work.

Microphone

Asked once, the first time you record. You can change it later in your system's privacy settings.

Updates

Shows your version. RevFly checks for new releases and installs them in one click.

02 · General

Hotkeys and microphone

The Input section controls how you start a recording and what RevFly listens with.

Launch hotkey

Starts and finishes a recording. When you finish, RevFly transcribes, translates if needed, and pastes. On a Mac, Right ⌥ works best: you almost never press it on its own, so it doesn't clash with other shortcuts. Press it once to start and again to stop.

Translate hotkey

Translation sets when RevFly translates: Off, Separate hotkey (the default) or Automatic. With a separate hotkey, the launch hotkey keeps the language you speak and the translate hotkey translates to your target language; the pill turns teal and shows the target, like EN. Automatic translates anything that isn't in the target language. Use another modifier key like Right ⌃ on a Mac, or a key combo on Windows and Linux. What you actually said is kept: open a translated recording with Second try in History to see and copy it.

Microphone input

Pick the device to record from. Test Mic plays it back so you can check the sound is clear and free of noise. If the mic drops out mid-recording, RevFly reconnects, or switches to the system mic if it was unplugged.

Audio output

The speaker for the start and finish chimes and for playing back recordings.

Noise reduction

Filters background noise (fans, traffic, keyboard) before recognition. Off, Light, Balanced or Strong. Try Light in a noisy room.

Open at login

In Behavior, below Input. RevFly starts when you log in, so the hotkey is always ready.

03 · Transcription

Speech model and languages

Recognition runs entirely on your computer. Pick the engine and tell it which languages to expect.

Transcription model

Parakeet TDT 0.6B is the default. It downloads on first launch and is very fast, with 25 languages. Whisper Base, Small and Medium download in one click; Medium is the most accurate.

Hardware acceleration

Auto uses your GPU or Neural Engine when there is one, which makes recognition faster and lighter. Without one, RevFly runs on the CPU alone.

Spoken language

Auto-detect by default. If you usually speak one language, select it: the model stops guessing, and results are more reliable.

Excluded languages

Languages you don't speak but that sound close to yours, for example Russian or Polish for a Ukrainian speaker. If one is detected by mistake, the speech is recognized again in your language.

Models folder & import

Open the folder where models are stored, import your own Whisper .bin file, or download one from a link.

04 · Transcription

Memory and long recordings

Choose how the speech model uses memory, and how long recordings are split so nothing is lost.

Idle timeout

How long the speech model stays in memory after you finish (30 s by default). After that it's unloaded, so RevFly uses almost no RAM while you're not dictating. It loads again automatically the next time you record.

Load speech model

Faster loads the model as soon as you start talking, so the text is ready the moment you stop. Save memory loads it after you stop and uses less RAM while you record.

Audio chunking

Not every model can take a long conversation in one go: Parakeet handles about 4 minutes 50 seconds at a time. Chunking splits the recording at natural pauses, so turn it on for long dictation. It's also faster: the parts you've already said are transcribed while you keep talking, so the text is ready almost as soon as you stop.

Pause detection trigger

How long a pause has to be to cut a chunk cleanly (500 ms by default).

Safety cut limit & overlap

If you talk without pausing, a chunk is cut after this many seconds. The overlap repeats a short slice so no word is cut in half.

05 · Text

Post-processing & formatting

Speech is messy. Normalization turns what you said into clean written text that reads well and that AI tools understand.

Fillers & stutters

Removes "um", "uh", and repeated words like "I I want".

Numbers & values

"Twenty four" becomes 24; currency, dates and percentages get a consistent format.

Self-corrections

"On Monday, no, Tuesday" becomes "On Tuesday".

Confidential masking

Further down, list names and codenames to hide. Emails, phone and card numbers are masked before text is translated or pasted.

06 · Translations

Translation

Speak your native language, or any language, and get the text in another one. Pick the target language; speech that's already in it is pasted without translating.

LLM, local or cloud

Presets for Gemini, OpenAI, Ollama and LM Studio. A free Gemini API key includes a number of free requests every day, and gemini-3.6-flash translates quickly and well. Ollama and LM Studio run on your computer, offline and without a key.

Prompt template

Tells the LLM what to do with your text. Besides translating, you can ask it to fix grammar or polish your phrasing.

Custom API

Any REST translation service. Several have free tiers, such as Google Cloud Translation.

If it fails

The pill shows the reason and a Retry translation button. The tray menu has Retry Last Translation too, so you never have to say it again.

07 · History

History and storage

Choose what RevFly keeps on your disk. Everything stays local.

Save format

Audio + Transcript, Transcript only, or Don't save (nothing is written to disk).

Storage cap & auto-purge

When the cap is reached, the oldest recordings are deleted first. Auto-purge removes anything older than the period you pick, or never.

Open folder

See your recordings as ordinary audio files.

Tip: keep audio on. A saved recording can be transcribed again later, as the next two sections show.

08 · History

Logs: see every step

Each recording gets a timeline of what happened to it and how long each step took.

Speed

Silence trimming, recognition, translation and paste times, plus how many times faster than real time the model ran.

Text at each stage

The detected language, the raw transcript, and the final text that was pasted.

Second try

Play the recording back, or open it in Audio Labs to transcribe it again. For a translated recording, it also shows what you said, ready to copy.

09 · Audio Labs

Audio Labs: test and re-run

A workbench for your real recordings. Pick a previous recording or drop in any audio file, then run it through different settings.

Compare settings

Try other models, languages, noise reduction and clean-up options, and compare the raw and cleaned text and the speed.

Recover a recording

Something went wrong, like the wrong language or a failed translation? Your recording is still there. Transcribe it again instead of repeating yourself.