The ONLINE / LOCAL switch
The switch at the top of the RECOGNITION panel decides where your speech is recognised. ONLINE sends it to a server and needs the internet. LOCAL runs a model on your computer, and the audio goes nowhere.
Each side remembers the model you last chose on it. Flip the switch and that model comes back. A side you never set runs its default: Google online (Soniox if you have Pro), and "Auto — by language" locally.
The model chooser
The Model plate below the switch shows the engine that runs now, with its Quality and Speed lamps (five each). Click it and the chooser slides out over the INPUT and POWER panels.
- On the left: every model of the current side, with its lamps and whether it is downloaded.
- On the right: the highlighted model's details. When its text appears, which of your languages it knows (✓ or —), when recording starts, whether it works without the internet, what it runs on, and all its languages.
- One button at the bottom does the action: Choose, Download and choose, Downloading…, ✓ Chosen, or About Pro for Soniox on the free plan.
Clicking a row only shows its details; the button makes it run. "◀ Hide" or Esc closes the chooser. A download keeps going after you close it.
The engines side by side
| Engine | Side | Text appears | Languages | Quality / Speed | Plan |
|---|---|---|---|---|---|
| Soniox | ONLINE | while you speak | finds the language itself, even mid-sentence; 33 in the app, 60+ in all | 5 / 5 | Pro |
| ONLINE | while you speak | one at a time; 11 languages | 3 / 5 | Free | |
| GigaAM v3 | LOCAL | after each phrase | Russian only, with punctuation | 5 / 4 | Free |
| T-one | LOCAL | while you speak | Russian only | 3 / 5 | Free |
| Parakeet v3 | LOCAL | after each phrase | 25 European languages, finds which one | 4 / 4 | Free |
| Whisper | LOCAL | after each phrase | 33 in the app, about 99 on AUTO; finds the language | 4 / 2 | Free |
| Vosk | LOCAL | while you speak | one language per model, 15 languages | 2 / 5 | Free |
On the free plan all engines share one allowance of 2,000 words a day; with Pro there is no daily limit. The full language list of each engine is on the languages page.
What each one is for
- Soniox is the most accurate and handles mixed languages. A recording can last as long as a meeting. It uses Pro credits, one per minute.
- Google is free live text through Google Chrome, one language at a time. It starts listening about a second after you press the hotkey, so speak after the quiet tick.
- GigaAM v3 (Sber) gives the best accuracy on natural Russian speech.
- T-one (T-Bank) is light and fast for Russian, with text while you speak.
- Parakeet v3 (NVIDIA) is accurate and fast for European languages, even without a graphics card.
- Whisper (OpenAI) covers the most languages. It is best with an NVIDIA graphics card; on the processor alone the app warns that it is very slow.
- Vosk runs on any computer, including old ones, but makes more mistakes.
Auto — by language
The first entry on the LOCAL side, marked RECOMMENDED, picks the best downloaded model for the language you dictate in:
- Russian: GigaAM v3, then T-one, then Whisper, then Vosk.
- A language Parakeet knows: Parakeet v3, then Whisper, then Vosk.
- Any other language: Whisper, then Vosk.
Where Whisper would run on the processor, Vosk goes ahead of it. Hover over the Model plate to see which model runs now. If no offline model is on the computer yet, a strip under the languages offers the one that suits your languages best, with a Download button.
When Soniox is not available
If Soniox cannot be reached, for example without the internet or after an error, dictation continues through Google at once with the note "Soniox is not available now — dictating through Google. Next time Soniox is tried again." Your choice of Soniox stays. When the month's credits are used up, dictation also goes through Google until they renew; see Soniox credits.