Whistle (offline)
Turns the speech in your video into timed subtitles on your own computer, with Whistle, the speech model from Cactus Compute. The 17 MB model comes inside the add-on: nothing to download on first use, no graphics card, and nothing leaves your machine.
Settings Add-ons Whistle (offline) Install
Subtitld downloads it, checks its SHA-256, and keeps it updated.
- Languages
- 7
- Models
- 1
- Add-on download
- 23–31 MB
- RAM, at least
- 512 MB
What it does
The model and its engine are inside the add-on. Install it, and it transcribes straight away, even offline.
It runs on the CPU and needs little memory. Twenty minutes of audio take about two minutes on a desktop computer.
Long audio is transcribed about a minute at a time, so the progress bar keeps moving and Cancel takes effect within a minute.
List the names and jargon your video uses under Keywords, and Whistle favours them when it is unsure of a word.
Models
There is one model, Whistle, and it is already inside the add-on.
| Model | Download | Good for |
|---|---|---|
| Whistle Default | Clear speech in its seven languages |
Whistle doesn't always add commas and full stops. Where it does, each subtitle holds one sentence; where it doesn't, subtitles are split at about two lines of text or 7 seconds.
Options
In Settings › Add-ons › Whistle (offline) › Configure.
Names and jargon from your video, separated by commas. When Whistle is unsure of a word, it leans towards these, so a name it would misspell comes out right.
Languages
English, German, French, Spanish, Italian, Dutch and Polish. Leave the language empty to detect it. A project in any other language is refused rather than guessed: use Whisper (offline) for those.
endefresitnlpl
Platforms
| System | Download | Tested |
|---|---|---|
| Linux x86-64 | 30.8 MB | On real machines and in CI |
| macOS (Apple silicon) | 28.7 MB | In CI |
| Windows x86-64 | 22.6 MB | In CI |
Before every release, each build transcribes a sample in CI, installed the way Subtitld installs it.
Troubleshooting
unsupported_languageWhistle doesn't know your project's language. Use Whisper (offline), which knows 99.model_missingA file is missing from the add-on. Reinstall it.bad_paramsSubtitld couldn't hand the audio over, or ffmpeg wasn't found. The message says which.oomOut of memory. Close other programs, or transcribe a range instead of the whole video.disk_fullNot enough disk space to decode the audio. Free some and try again.
License
Fine for commercial work: the add-on, the Needle engine and the Whistle weights are all Apache-2.0.
Versions
- 26.10.1
Long audio no longer looks stuck at 20%.
Whistle used to transcribe the whole file in one go: the progress bar stopped at 20% until it finished, Cancel had to wait for the end, and long files slowed down more than their length (20 minutes took about 3 minutes). The audio is now transcribed about a minute at a time, cut at quiet moments so no word is split:
- The progress bar moves after every minute of audio.
- Cancel takes effect within a minute.
- 20 minutes of audio take about 2 minutes instead of 3, with the same words.
- With no language set, the first stretch of speech sets the language for the rest of the file.
Versions now follow Subtitld's own: year, month and the release of that month. 26.10.1 comes after 0.0.2, and Subtitld offers it as an update.
The model, the engine and the add-on are unchanged otherwise: Apache-2.0, offline, CPU only.
Other Speech to text add-ons
-
Whisper Offline
Offline speech to text with whisper.cpp, CPU only.
-
Vosk Offline
Light and CPU-only. Models from about 40 MB.
-
RealtimeSTT Offline
For Record mode, as you speak. Whisper models, on your computer.
-
AssemblyAI Cloud
With speaker diarization. Not available yet.
-
Phonon-2 Offline
Offline English speech-to-text with Phonon-2 (Fermion Research): a 164 MB model that runs on the CPU, downloaded…