Whisper (offline)
Turns the speech in your video into timed subtitles, on your own computer, with whisper.cpp. No cloud account, no graphics card, no Python to install. Pick a model once; after its download, nothing leaves your machine.
Settings Add-ons Whisper (offline) Install
Subtitld downloads it, checks its SHA-256, and keeps it updated.
- Languages
- 99
- Models
- 9
- Add-on download
- 21–46 MB
- RAM, at least
- 512 MB
What it does
Segments appear on the timeline while it decodes, then are regrouped so each subtitle holds one sentence instead of a fragment.
Transcribe the whole video, a stretch of it, or only the subtitles you selected.
The model stays loaded between phrases, so only the first one waits for it to load.
Stops within about a second, even mid-decode, and never leaves half a result behind.
Upgrading from 26.03? This is the engine that used to be built in. It keeps the same id, settings and model files, so your setup and downloaded models carry over without a new download.
Models
Bigger models make fewer mistakes and take longer. Base is the default and a good start; try Small or Medium for names, accents and noisy audio. The .en versions understand only English and are a little faster.
| Model | Download | RAM | Good for |
|---|---|---|---|
| Tiny tiny · tiny.en | ~0.3 GB | Quick drafts, slow computers | |
| Base Default base · base.en | ~0.4 GB | Clear speech, most videos | |
| Small small · small.en | ~0.9 GB | Accents, names, interviews | |
| Medium medium · medium.en | ~2.1 GB | Noisy audio, several languages | |
| Large v3 large-v3 | ~3.9 GB | The best result, when time allows |
Models come from ggerganov/whisper.cpp on Hugging Face, once, the first time you use them.
Where models are kept
In Subtitld's shared model cache, never inside the add-on, so updates don't delete them. Uninstalling the add-on keeps them too; delete the folder yourself to free the space.
| System | Folder |
|---|---|
| Linux | ~/.cache/subtitld/models/whispercpp |
| macOS | ~/Library/Caches/subtitld/models/whispercpp |
| Windows | %LOCALAPPDATA%\subtitld\subtitld\Cache\models\whispercpp |
Options
In Settings › Add-ons › Whisper (offline) › Configure.
baseWhich model to use. See the table above.
0Auto uses up to 4 threads. Raise it on a computer with more cores to go faster.
One subtitle per sentence. Turn it off to keep Whisper's own segments, which often break mid-sentence.
Empty shares Subtitld's model cache. Point it elsewhere to keep models on another drive.
Languages
99 languages, or leave the language empty to detect it. Regional tags such as pt-br use their main language; every Chinese variant maps to zh. Of Subtitld's languages, only Zulu isn't supported.
afamarasazbabebgbnbobrbscacscydadeeleneseteufafifofrglguhahawhehihrhthuhyidisitjajwkakkkmknkolalblnloltlvmgmimkmlmnmrmsmtmynenlnnnoocpaplpsptrorusasdsiskslsnsosqsrsusvswtatetgthtktltrttukuruzviyiyozh
Platforms
| System | Download | Tested |
|---|---|---|
| Linux x86-64 | 46.3 MB | On real machines and in CI |
| macOS (Apple silicon) | 21.4 MB | In CI |
| Windows x86-64 | 21.7 MB | In CI |
Before every release, each build runs a real transcription in CI, installed the way Subtitld installs it.
Troubleshooting
model_missingThe model file is missing or damaged. The message names the file; delete it and run again to download it fresh.network_unavailableThe model couldn't download. Try again; behind a proxy that inspects HTTPS, setSSL_CERT_FILEto its certificate.disk_fullNot enough space for the model. Free some, or move the models folder.oomOut of memory. Pick a smaller model.unsupported_languageWhisper doesn't know this language. Try Vosk, or auto-detect.
License
Fine for commercial work. Some other add-ons use models for non-commercial use only; each page says so at the top.
Versions
- 0.0.1
Offline speech-to-text for Subtitld with whisper.cpp (pywhispercpp 1.5.1).
Models are not included: the selected GGML model downloads once on first use into Subtitld's shared model cache, so models downloaded by earlier Subtitld versions are reused.
The add-on is Apache-2.0. Each archive contains
THIRD_PARTY_LICENSES/with the licences of the third-party components frozen into it: whisper.cpp and pywhispercpp are MIT, pybind11 and numpy are BSD-3-Clause, and the Python runtime and its libraries are under their own permissive licences.The Linux archive also ships, inside numpy's OpenBLAS, the GCC Fortran runtime (GPL-3.0 with the GCC Runtime Library Exception) and libquadmath (LGPL-2.1-or-later). The complete source of that libquadmath build is attached here as
whispercpp-addon-<version>-libquadmath-source.tar.gz; the add-on's own source iswhispercpp-addon-<version>-source.tar.gz.