Whisper, natively

Whisper subtitles on your Mac, without the terminal.

whisper.cpp with Metal acceleration, Large v3 Turbo, Silero VAD and word timing measured at about 3 ms — wrapped in an editor, so the three words it mishears take ten seconds to fix instead of ten minutes.

Get the Mac app — $9.99 → Try it free in your browser
Whisper without the terminal

The good part of Whisper, minus the setup.

If you have already run whisper.cpp from a shell, you know it produces a usable transcript and a fairly unusable subtitle file.

Whisper is very good at hearing words and fairly indifferent about when they were said. Run it from the command line and you get an SRT with sensible sentences and timings that drift at phrase boundaries — close enough for a transcript, visibly late for a caption that is meant to land on the beat.

suubbs runs the same engine and then does the work afterwards that the CLI leaves to you: it aligns word boundaries against the audio, segments lines against real reading-speed rules, and gives you somewhere to fix the three words it got wrong without hand-editing timestamps in a text file.

Words are the source of truth

This is the design decision everything else rests on. A cue's text is always derived from its words and never stored beside them. That is what makes word-level highlighting, retiming, and re-segmenting with a better model while keeping your edits possible at all. Most subtitle tools store the line and throw the word timings away, which is why re-segmenting in them means starting over. The same per-word times can be exported as data, too — word-level timestamps shows the file.

1 · What happens when you press Generate

Whisper is one step of six.

The model hears the words. Most of what makes the result a usable subtitle file happens before and after it.

Run whisper.cpp yourself and you are responsible for everything around the model. In suubbs' Mac app, the full pipeline runs on your machine in this order:

  1. Decode the audio in-process. The video's soundtrack is demuxed and decoded inside the app — no separate ffmpeg step, and no file written anywhere you have to clean up.
  2. Find the speech. Silero VAD, a voice-activity model of about 1.5 MB, marks where someone is actually talking. It is set to treat a stretch as speech only if it lasts at least 250 ms and to split at silences of 100 ms or more, padding each speech region by 60 ms so first and last syllables are not clipped. Silence never reaches the model, which is the main defence against Whisper's habit of inventing text over a quiet room tone.
  3. Transcribe, one word at a time. whisper.cpp runs with Metal acceleration on the GPU, with greedy decoding and its output split into one word per segment, so every word carries its own timestamp rather than inheriting a phrase's. A no-speech threshold and blank suppression are a second guard against hallucinated lines, independent of the VAD.
  4. Align each word. whisper.cpp's dynamic-time-warping timestamps are computed from cross-attention using alignment heads that must match the exact checkpoint. suubbs picks the heads from the model's filename, and if it cannot identify the model it switches DTW off rather than guess with the wrong heads.
  5. Correct the timing against the waveform. Neither of Whisper's timings is good enough on its own, and they fail in opposite directions: segment times run about 240 ms early after a pause, and DTW times trail the true onset by 70–240 ms. suubbs looks for real pauses in the audio — at least 70 ms of genuine quiet, which is longer than the silent closure inside a consonant — and moves each boundary onto the pause's edge, by at most 250 ms. Inside continuous speech, where no pause exists, it leaves the model's estimate alone. Why raw word times run late is explained in more depth on word-level timestamps.
  6. Build readable lines. The word stream is cut into subtitles against real rules: 42 characters and two lines for widescreen (22 and three for vertical), at most 20 characters per second of reading speed, each subtitle on screen for between 0.833 and 7 seconds, and a new subtitle forced at any silence over 0.6 seconds. The rules are explained on subtitle reading speed.

Each word also keeps its model confidence. The editor marks words whose timing was interpolated rather than aligned, and lines that read too fast or run over the line width — none of which blocks an export, but all of which tell you where to look first.

DTW alignment, the waveform correction and Silero voice detection are each a switch in the transcribe settings, all on by default, so you can turn one off and hear the difference for yourself.

2 · Choosing a model

Three Whisper models, downloaded once.

The app does not bundle models. You download the one you want on first use, and from then on it works offline.

Model Download Languages Good for
Base (English)148 MBEnglish onlySpeed. Comparable to what the browser build runs.
Small (English)488 MBEnglish onlyGood accuracy per megabyte.
Large v3 Turbo, quantised574 MBMultilingualThe recommended default: best accuracy for the size.

Sizes here are decimal megabytes, the way macOS Finder counts them. The model menu in the app's Transcribe sheet divides by 1,048,576 instead, so it lists the same three files as about 141, 465 and 547 MB — the same downloads, counted in binary units.

The Silero VAD model is fetched alongside whichever speech model you choose. Models are kept outside the app bundle, so updating suubbs does not download them again. The English-only models cannot be asked to transcribe another language, and the app disables the language picker when one is selected; for anything other than English, use Large v3 Turbo.

Translate to English. Whisper can translate as it transcribes, and suubbs exposes that as one switch. It is available with the multilingual model whenever the spoken language is not English, and produces English subtitles timed to the original speech.

In the browser. The free trial at suubbs.com/app is the same studio running Whisper through Transformers.js instead of whisper.cpp. It transcribes the first 60 seconds, defaults to a smaller model (Base, about 78 MB), has no Silero VAD — it uses a cheaper silence stand-in — and cannot take a names-and-jargon prompt, which is why its transcripts and word timings are rougher than the Mac app's. Everything after transcription — editing, styling, export — is the same code.

If a transcript comes back with wrong words, improving subtitle accuracy goes through the four usual causes and the setting that fixes each one.

All six steps, one button.

[Generate]

Against the command line

What you give up by running Whisper yourself.

Nothing here is impossible with the CLI. It is a question of how many evenings you want to spend on it.

  whisper.cpp by hand suubbs
SetupHomebrew, a compiler, ffmpeg, model files fetched by handDrag one app to Applications
Audio extractionAn ffmpeg incantation per fileDemuxed and decoded in-process
Word timingRaw DTW output, late at phrase boundariesDTW snapped onto real amplitude edges, about 3 ms
Fixing a wrong wordReopen the SRT in a text editor and re-count timestampsClick the word, type, timings hold
Re-segmenting linesRerun with different flags and hopeWords are the source of truth, so lines re-flow without losing edits
StylingHand-write ASS override tagsLive preview that matches what libass renders
CostFree$9.99 once
Under the hood

For people who will ask.

The engine, the models, and the honest limits.

Which Whisper model does it use?

Large v3 Turbo by default on the Mac app — 574 MB, downloaded once on first launch. The browser version starts with a 78 MB model, which is why its transcripts are rougher.

Is it whisper.cpp or the Python version?

whisper.cpp, with Metal acceleration, wrapped by a Rust core. There is no Python runtime anywhere in the app.

How accurate is the word timing?

About 3 ms against known speech onsets on the desktop build. Cross-attention alignment gives the initial boundaries and an acoustic pass snaps them onto genuine amplitude dips. Inside continuous speech, where no pause exists, it deliberately leaves the boundary alone rather than inventing one.

Can I bias it toward names and jargon?

Yes, on the Mac app. Give it the proper nouns before you start and the transcript uses your spelling instead of guessing. The browser build cannot do this — Transformers.js declares the prompt parameter but never consumes it.

Does it do voice activity detection?

Silero VAD runs in front of the model on the desktop build, which is what stops long silences being transcribed as hallucinated speech.

Do I need ffmpeg or Homebrew installed?

No. The Mac app decodes the audio itself and ships whisper.cpp inside the app. Nothing has to be installed from a terminal, and no Python runtime is involved.

Can I get Whisper's word timestamps out as data?

Yes. The Word-level JSON export keeps every word with its start, end and confidence, after the waveform correction. SRT, WebVTT and ASS exports are built from the same words.

Does it work without an internet connection?

Once the model has been downloaded, yes. Transcription, editing and export all run on the Mac. The network is used to fetch a model the first time, to check for an app update at launch, and to fetch a translation model if you translate into a new language.

Which languages?

Sixteen, detected automatically: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Japanese, Korean, Chinese, Hindi, Arabic, Turkish and Ukrainian. Any of them can also be translated into English.

Keep reading

Related

Same engine. Someone else did the plumbing.

[no homebrew required]