Transcription
Transcription gives every word of the video a start and an end time. You decide three things for each video, in this order: where it happens, with which speech model, and in which language. The web app and the iPhone app ask the same two questions: first where and with which model, then the spoken language.
Where it happens
Section titled “Where it happens”| Provider | Where the audio goes | Use it when |
|---|---|---|
| Local (the default) | Nowhere. A model runs on the server’s own machine. | You want privacy and have a few minutes of CPU, or a GPU. |
| Another OpenCaptions server | To the server you name, over its API. | Your laptop has no GPU and a machine of yours has one. |
| OpenAI | To the OpenAI API, with your key. | You do not want to run a model at all. Their data policy applies. |
The page says so whenever a provider that sends audio away is selected. The two remote providers are set up by the person who runs the server: see Speech models and providers.
On the iPhone, the default is the phone itself, and a server of yours is an option in the settings. See The iPhone app.
Speech models
Section titled “Speech models”| Model | Languages | Good for |
|---|---|---|
| Whisper Large v3 Turbo (default) | About 100 | The best all-round accuracy. About 1.6 GB on a server, 627 MB on the iPhone. |
| Whisper Large v3 | About 100 | The most accurate Whisper, and slow on a CPU. About 3 GB. |
| Whisper Small | About 100 | Smaller and faster, less accurate. About 480 MB, and a fine first try. |
| Parakeet v3 | 25 European languages | Fast: about 2.5 times quicker than Large v3 Turbo on a CPU, a little less accurate. About 0.65 GB. |
| Apple Speech | Those iOS offers | The iPhone only, built into iOS 26, with nothing to download. |
The web app and the iPhone app offer the same models. On a server, the first use of a model downloads it.
Parakeet v3 covers Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. If a recording is in another language, the server hands it to Whisper and the job says so.
Accuracy varies by language, and the numbers are in the speech model benchmarks. In short: Whisper Large v3 Turbo is the safe choice; Parakeet is the one to pick when speed matters and your language is on its list.
The spoken language
Section titled “The spoken language”Pick the language the video is spoken in from the searchable list, which has a flag for each. The list follows the model: Parakeet shows its 25, Whisper shows them all. The language you used last is selected to start with, otherwise your browser’s or phone’s.
There is no “detect it” choice in the apps: the language is always chosen. (The API still accepts
auto for scripts.)
Captions can be drawn in Latin, Cyrillic, Arabic, Hebrew, Indic, Thai and Chinese, Japanese and Korean scripts. Words in a script the chosen font lacks are drawn with a fallback font.
What it does with silence and noise
Section titled “What it does with silence and noise”The speech models can invent words over music or noise, or skip a loud stretch altogether. OpenCaptions decodes the whole recording rather than letting a voice detector cut it, and cuts a line off at a long silence so invented text does not run on. On the iPhone, a second pass re-reads the stretches where nothing was heard but there is sound, and keeps only what two readings agree on. The benchmarks explain why.
If a transcript comes back empty, the app says that no speech was found.
Transcribe again, or bring your own
Section titled “Transcribe again, or bring your own”- Re-transcribe (in the editor’s toolbar) runs a project again with another model or language. It replaces the transcript, so your word edits are discarded.
- Edit as text opens the transcript as JSON, and imports a subtitle file (SRT, WebVTT, with or without per-word times, or OpenCaptions JSON). See Editing.