Frequently asked questions

Yes. Speech recognition runs locally on your computer — you only need internet once, to download a model. Cloud engines (OpenAI, Yandex SpeechKit) are optional; the default is fully offline.

Whisper recognizes 90+ languages, NVIDIA Parakeet covers 25 European languages with built-in punctuation, and GigaAM is a specialist model for Russian. Pick the one that fits how you speak. The app itself is available in English and Russian.

Very accurate. Whisper large-v3-turbo handles everyday speech, jargon and mixed-language sentences well, and Parakeet is faster still on the languages it covers. Punctuation and capitalization are added automatically, and a custom dictionary fixes names and terms the model gets wrong.

Built-in dictation sends your voice to Microsoft’s cloud and needs internet for most languages. SweetWhisper works locally and offline, lets you choose the model, inserts text into any app and can clean up the result with AI (Pro).

Speed (no network latency), privacy (audio never leaves your machine), cost (no per-minute fees) and it works anywhere — even on a plane.

No. Dictation is free, with no limits, forever. Pro is a one-time purchase that you keep for good: it adds AI cleanup, translation, task modes and more. Try it first: 14 days free on Windows, or your first 200 floating-button dictations on Android.

Yes. On Windows, every Pro feature is free for 14 days after you install. On Android there’s no time limit: you can use the Pro features — the floating button, translation and the largest models — for your first 200 floating-button dictations, and the app shows how many are left. No card, no sign-up. When the trial ends, Pro features lock until you add a key; regular dictation, including dictating from the keyboard on Android, stays free with no limits. One key unlocks Pro on both your computer and your phone.

No — unless you turn that on yourself. By default Pro uses the built-in local AI: text is processed on your PC and never leaves it, which is safe even for confidential documents. Cloud models are connected only manually, with your own key.

The key arrives by email as soon as the payment goes through — it is issued automatically. Open “Settings → Pro”, paste it and hit “Activate” — Pro unlocks instantly, no internet required. Step-by-step guide with a screenshot: activating your Pro key. Nothing after ten minutes? Check your spam folder, then write to hello@sweetwhisper.app.

Yes. One key covers 3 of your personal devices — just paste the same key on the new PC. It is verified offline, with no account and no internet check. Reinstalling Windows or replacing your drive won’t use up the key: activate it again the same way.

Nowhere. Audio is processed on your PC and never leaves it. Transcription history lives in a local database you can clear at any time.

Everything is local. The portable edition keeps it in the folder next to the exe; the installed edition uses %AppData%\SweetWhisper (history, recordings, settings) and your models folder. You can clear the transcription history from inside the app, and to wipe your data just delete that folder. One thing is kept separately: an anonymous trial marker (in the Windows registry and %LocalAppData%\SweetWhisper) so the 14 free Pro days are granted once per device. It holds none of your data, is never sent anywhere, and you can delete it by hand if you want. Nothing is left on our servers, because nothing is ever sent there.

No. Audio is processed on the device and discarded; the transcription history lives in the phone’s storage only, and you can switch it off or wipe it. Android backup is disabled, so history never reaches Google’s cloud either. There is no account, no ads, no analytics and no trackers, and password fields are never recorded.

Windows 10/11, a graphics card is optional: with one recognition is faster (any NVIDIA, AMD or Intel works), without one everything runs on the CPU. Recognition only runs while you dictate — the rest of the time the app just waits for the hotkey. Models take 500 MB – 3 GB of disk space.

Linux support is in the works. The core already runs; a full release will follow.

Yes. SmartScreen shows that prompt for files few people have downloaded yet: reputation grows with each install, and every new version of an unsigned app starts from scratch. It warns about novelty, not malware. Click “More info” → “Run anyway”. To confirm the file is really ours, check its checksum — we publish them with every release, linked from the download section.

Downloads resume from where they left off — just click “Download” again, nothing already fetched is lost. If huggingface.co is blocked on your network, the app automatically falls back to a mirror; you can also turn on a VPN for the download. Models are large (500 MB to 3 GB), so on a slow connection this takes a while.

Check that the right microphone is selected in Settings → Audio, and that Windows lets apps use the microphone (Windows Settings → Privacy → Microphone). On startup the app checks the mic and tells you if it isn’t ready. If you have several microphones, pick the correct one from the device list.

The recording key is set in Settings → General → Hotkeys: click “Record combination” and press your own. If the combination is already used by another app or switches your keyboard layout, the OS drops the held keys and recording is cut off — choose a combo with a regular key (for example Ctrl + Alt + V) and avoid combinations with the Win key.

Drag the bar with the mouse to any corner of the screen. To hide it or show it again, use the tray icon menu (the SweetWhisper icon near the clock). In fullscreen games and video the bar stays hidden so it never gets in the way.

Google sends your speech to its servers; Sweet Whisper recognizes it on the phone itself, so it works on a plane or on the subway, collects nothing and shows no ads. It doesn’t replace your keyboard either: it steps in when you tap the microphone and hands you straight back.

Only once, to download a recognition model. After that everything happens on the phone: you can dictate with no connection at all, and nothing is ever uploaded.

The app itself is about 36 MB, plus one model: 41–388 MB for a compact Whisper model, 670 MB for Parakeet (25 languages), or up to 874 MB for a full Whisper model (better on newer phones). GigaAM, a Russian-only model, is 232 MB. Models are downloaded one at a time, can be deleted, and the catalog shows how much space is in use. On first run the app suggests a model that fits your phone’s memory and core count.

A Play release is being prepared; for now we serve the APK from our own site. It is signed with our key, so updates install over it and no one can swap in a different build. Android will ask once whether you trust the source you are installing from — that is the normal flow for apps outside the store. Download it only from sweetwhisper.app; we publish no copies anywhere else.

Yes, under two conditions. Only the Whisper models can translate, and only into English — with GigaAM (Russian) or Parakeet (25 languages) the “Translate” switch is disabled, they transcribe very accurately but never translate. Translation is a Pro feature: try it free during your first 200 floating-button dictations; after that it needs a Pro key — the same one you use on Windows.

Yes — that is server mode (the Corp business tier). The app on the strong machine accepts dictation from other devices on your network and transcribes it on its own GPU, so a weak laptop needs no model of its own. Access is by key only, up to three keys, each revocable in one click. It listens only on that computer and is reached through a secure tunnel, so no ports are opened. Other people’s speech never lands in your history or on your disk, and your own dictation always takes priority.