Most wrong words in automatic subtitles come from four places: a speech model too small for the audio, the wrong spoken language set, names and jargon the model has never heard, and silence or music it fills with invented text. Fix those before you start correcting, in that order — a bigger model, the right language, a list of the names, clean audio — and the corrections left to do by hand shrink to a few lines. In suubbs the model, the language, the names list and voice detection are all in ⋯ → Transcription settings, and they only help if you set them before you press Generate.
Every common kind of mistake follows from that.
suubbs uses Whisper, OpenAI's open-source speech model. Whisper does not look words up; it predicts the text most likely to go with the sound, one token at a time. When the sound is clear and the words are ordinary, the likeliest text is what was said. When it is not, the model still produces a fluent sentence — which is why wrong captions so often read as plausible English that nobody said, rather than as obvious garbage.
That gives four distinct failures, and they need different fixes:
All of them are in ⋯ → Transcription settings.
Set these first, then correct. Pressing Regenerate transcribes the file again from scratch and replaces the whole transcript — your hand corrections and the undo history with it. Get the settings right on a first pass, and only then start fixing individual words.
A model can only transcribe what reached the microphone.
If you can still re-record or re-export, these change accuracy more than any setting does:
Timing is a separate question from accuracy. On the Mac, word boundaries are aligned and then snapped onto real pauses in the waveform, measured at about 3 ms from true speech onsets; in the browser they land within roughly 60–140 ms. Either is fine for ordinary subtitles, where whole lines come and go. It matters for word-by-word captions, and the word-level timestamps page explains how it is measured.
What happens to the timing depends on how you edit.
Click a line in the cue list and retype it. When you finish, suubbs lines your new words up with the old ones position by position:
For ordinary subtitles this does not show: a line still starts and ends where it did. It shows in a word-by-word highlight, where an estimated word can light up slightly early or late. The ⋯ menu tells you how many words have interpolated timing, so you know whether there is anything to check, and none of it blocks an export.
suubbs does not mark the words it was unsure of, so the reliable check is to play the video and read along. Look hardest at the places the four failures above concentrate: names and numbers, the first line after a long pause, anything over music, and the passages where people talk over each other. A line flagged as reading too fast usually needs splitting into two, or rewording — the subtitle reading speed page explains the limit. If a line was invented outright over silence, delete it.
If your captions are right but appear early or late, that is a sync problem rather than an accuracy one; fixing subtitle sync covers it.
Same studio, two different speech engines underneath.
| Browser studio (free) | Mac app | |
|---|---|---|
| Default model | Base | Large v3 Turbo (quantised) |
| Default language | English | Detect automatically |
| Names and jargon list | No | Yes |
| Silence handling | Skips near-silent blocks | Silero voice detection |
| Beam search decoding | Optional | No |
| Word timing | Roughly 60–140 ms | About 3 ms |
| How much it transcribes | The first 60 seconds | The whole file |
The browser studio is the honest way to judge accuracy on your own footage before paying: the first 60 seconds are transcribed with the same studio, and every editing and export feature works on them. Both run on your own computer; neither uploads the video.
Mostly about names.
Because the model writes the likeliest text for the sound, and a rare name is less likely than common words that sound like it. In the suubbs Mac app, type the name into Names and jargon in Transcription settings before generating; it biases the transcript toward your spelling.
No. suubbs does not fine-tune the model. The Names and jargon list is the supported way to steer spellings, and it applies to the run you set it for.
The spoken language is set to English — the browser studio's default — or, in the browser studio, Translate to English is switched on. Set Spoken language to the language actually spoken, or to Detect automatically in the Mac app, and generate again.
Whisper sometimes writes text into silence or music. The Mac app runs Silero voice detection to prevent it, and the browser studio skips near-silent audio; anything that still gets through can be deleted from the cue list.
Only for the word you changed. Words you leave alone keep their measured timing; a replaced word gets an estimated time within its line, and adding or removing a word re-estimates the words after it in that line.
No. Regenerate transcribes the file again and replaces the whole transcript, including your edits and the undo history. Change the settings first, then correct.
By default, yes: it runs Large v3 Turbo where the browser starts on the much smaller Base model, and it adds the names list and voice detection. The browser can load Large v3 Turbo too, on a machine with enough GPU memory.
Not currently. Read the transcript through while it plays, concentrating on names, numbers, music and crosstalk.