suubbs for Mac transcribes on your own computer: Whisper Large v3 Turbo, running through whisper.cpp on the Mac's GPU. Download the speech model once — 574 MB — and from then on you can open a video, generate subtitles, edit them and export SRT, VTT or ASS with the wifi off. Nothing is uploaded, nothing is metered, and there is no copy of your video on anybody's server.
Not encrypted in transit. Not deleted after 24 hours. Not sent at all.
Almost every subtitle tool on the web works the same way: you upload your video, their servers transcribe it, and you download the result. The privacy page then explains how carefully they look after the copy they now have.
suubbs does the transcribing on your own machine. There is no copy anywhere else, so there is no retention window to trust, no breach that can expose your footage and no terms of service that can change next year.
"Offline" should be stated precisely, so here is the complete list of what the Mac app fetches. None of it carries anything of yours.
Those sizes are decimal megabytes, as Finder counts them. The model menu in the app's Transcribe sheet counts in binary units, so it lists the same three speech models as about 547, 141 and 465 MB.
The models are kept in ~/Library/Application Support/Suubbs/models. Once they are
there, transcription needs nothing from the internet.
The complete list of what suubbs ever receives about you is your email address, handed over by Stripe at the checkout, and the fact that you bought a copy. The app sends no analytics and makes no licence check — the app you download is the paid product, so there is nothing for it to ask permission about.
Five minutes on a good connection, before the flight rather than during it.
From then on the cycle — open, generate, correct, style, export — never needs a network. Generated subtitles can be exported as SubRip (.srt), unstyled SubRip for an editor to style, WebVTT (.vtt), Advanced SubStation (.ass) for burning in, or word-level JSON.
The same family of model the cloud tools use. The difference is where it executes.
The Mac app runs Whisper Large v3 Turbo through whisper.cpp, accelerated on
Apple's Metal, with Silero voice-activity detection in front of it so silence is not
transcribed as hallucinated speech. Word boundaries come from cross-attention (DTW) alignment
and are then snapped onto real pauses in the waveform, which measures at about
3 ms mean error against known speech onsets, against 143 ms for the
alignment alone. If you know the names and jargon in the recording, type them into the
"Names and jargon" field and the transcript is biased toward your spellings.
Everything it needs is inside the app. No Homebrew, no Python environment, no ffmpeg to install separately — audio is demuxed and decoded in-process. It reads MP4, MOV, MKV, WebM, WAV, MP3, FLAC and Ogg, with AAC, ALAC or PCM audio inside them. Less common audio — AC-3 from some cameras, broadcast formats — will not decode on its own; if an ffmpeg happens to be installed on the Mac, the app falls back to it, and if not, the error names the format it could not read. You download one file and it works, natively, with no Wine layer or Windows VM of the kind Subtitle Edit needs on a Mac.
Plenty of people caption cat videos and none of this matters. It matters when the footage is someone else's.
If your work is more transcription than subtitling — meeting notes, a podcast archive — a dedicated transcription app may suit you better; the suubbs vs MacWhisper page is honest about where that line falls. suubbs is built around the subtitle itself: cue timing, line breaks, styling and the export.
Free, private, and limited to a minute.
The free browser studio runs speech recognition in the tab through Transformers.js on WebGPU, so your video is never uploaded there either. It is not quite the same pipeline: it uses smaller models by default (Base, about 78 MB), stands in for voice detection with a simpler energy gate, has no names-and-jargon prompt, and its word timings land within roughly 60–140 ms rather than 3 ms. It captions the first 60 seconds of a file.
The model is cached by the browser after the first download, so it does not have to be fetched again. But a web page has to be loaded from the web, so the browser version is "nothing uploaded" rather than "no connection". For genuinely offline work, the Mac app is the one.
Answered including the parts that are limitations.
Yes, once the models are downloaded. The speech model is a one-time 574 MB download into ~/Library/Application Support/Suubbs/models, and the voice-detection model is about 1.5 MB. After that you can open, transcribe, edit and export on a plane.
To fetch the models the first time, to fetch a translation model if you translate into a new language, and to check for an app update at launch. That is the app pulling things down; nothing of yours goes the other way. Without a connection the update check fails quietly and the app carries on.
No. Nothing is metered, because nothing runs on our hardware. A ten-second clip and a two-hour interview cost the same: the one-time price of the app.
MP4, MOV, MKV, WebM, WAV, MP3, FLAC and Ogg, with AAC, ALAC or PCM audio inside. It decodes them itself, so nothing else needs installing. Rarer audio such as AC-3 from some cameras will not decode unless an ffmpeg is already installed on the Mac, in which case the app uses it.
A cloud tool needs a copy of your video on its servers to transcribe it. That copy is subject to its retention policy, its breach exposure and its terms of service. suubbs never receives the file, so none of those questions apply.
That is the case this is built for. If your contract says the footage does not go to third parties, a cloud captioning tool is a third party. A program running on your own machine is not.
It never uploads your video, and the browser caches the speech model after the first download. But a web page has to be loaded from the web, and it captions only the first 60 seconds of a file. For work with no connection, use the Mac app.
On the desktop, an outbound firewall app will show you each connection it makes: the model downloads and the update check, and nothing carrying your files. Easier still, open the free browser version and watch the network inspector — a tab cannot hide traffic from you.