Where a download starts
- In the model chooser: highlight a LOCAL model and press "Download and choose". The row and the details show the percentage; when it finishes, the model runs.
- Under the LANGUAGES row: when the chosen model is not on the computer yet (or, for "Auto — by language", no model at all), a strip says "Download needed:" with its name, size and a Download button.
- On a Vosk language key, which downloads that language's model.
Closing the chooser does not stop a download. Until a model is on disk, LOCAL runs the best one you already have.
Sizes
| Model | Download |
|---|---|
| T-one | about 144 MB |
| GigaAM v3 | about 232 MB |
| Parakeet v3 | about 670 MB |
| Vosk | one model per language, 31 to 133 MB; most about 40 to 50 MB, Ukrainian is the largest |
| Whisper | 75 MB to 1,031 MB, depending on the size you pick |
Whisper sizes
Whisper comes in five sizes. In the chooser, pick the size in Whisper's details (Tiny, Base, Small, Turbo or Large) before you download it; the FILES screen lists the same sizes under longer names. Bigger means fewer mistakes and slower work.
| Size | Name on FILES | Download | Good for |
|---|---|---|---|
| Tiny | Instant · Tiny | 75 MB | checking that everything works; clear English |
| Base | Fast · Base | 142 MB | simple recordings with clean sound |
| Small | Balanced · Small | 466 MB | a computer without a graphics card |
| Turbo | Best · Large v3 Turbo | 547 MB | the recommended size: nearly the accuracy of Large, several times faster |
| Large | Most accurate · Large v3 | 1,031 MB | noise, accents, several voices; slow |
Dictation and the FILES screen use the same Whisper files, so a size downloaded for one is ready for the other.
Whisper and the graphics card
Whisper is the one model that cares about your hardware. The chooser says what it runs on:
- With an NVIDIA graphics card, it offers a CUDA module: "Download CUDA". With it, Whisper runs ten times faster or more.
- Without an NVIDIA card, it warns: "No NVIDIA graphics card found: Whisper will run on the processor — very slow, in any language." For offline dictation on such a computer, Parakeet v3, GigaAM v3, T-one or Vosk are better choices.
The other offline models run on the processor and need no graphics card.
Where models are stored
Everything goes into the VoiceKeyboard folder inside your Windows profile's %AppData%: Whisper files in models, the other dictation models in models\live. Nothing is sent anywhere while a model works.
If a download fails
The row shows "Download failed:" with the reason. Check the internet connection and press the download button again. If a model on disk will not start, the app says "Could not start the local model. Try downloading it again."