WinSTTFAQ
Direct answers about WinSTT — privacy, offline use, languages, cost, GPU requirements, models, and where your data lives.
Short, self-contained answers to the questions people ask most before and after installing WinSTT.
Privacy & offline
Is my audio uploaded anywhere?
No. Transcription runs entirely on your machine. Audio frames are processed in memory by on-device ONNX models and discarded immediately — never written to disk, never uploaded. There are no usage analytics. The only outbound traffic from a default install is: anonymized crash reports (Sentry, opt-out), first-time model downloads from huggingface.co (only when you pick a model that isn't cached), and optional cloud LLM cleanup or cloud STT (only if you explicitly add an API key).
Does WinSTT work fully offline?
Yes. The base model is vendored with the app, so a fresh install transcribes with zero network access. You only need a connection to download additional models or to use opt-in cloud features (cloud LLM cleanup, cloud STT). Local LLM cleanup via Ollama also runs on-device, so even that path keeps your text on your machine.
Are my transcriptions stored on disk?
Only the optional History dashboard persists them, and it stores to a local SQLite database — nothing leaves your machine. Recordings are saved locally too, governed by a retention setting. Set WINSTT_LOG_NO_TEXT=1 to redact transcribed text from the rolling debug.log.
Your voice stays on your machine
Transcription is local-first by design. Audio never leaves your hardware, there's no telemetry, and the entire codebase is MIT-licensed and auditable.
Cost & licensing
Is WinSTT free?
Yes. WinSTT is free and open source under the MIT license. There are no accounts, no subscriptions, and no paywalled features. You can read, audit, fork, and modify every line.
Do I have to use any cloud services?
No. Cloud LLM cleanup (OpenRouter) and cloud STT (ElevenLabs, OpenRouter) are strictly opt-in and locked until you add your own API key under Integrations. With no keys configured, WinSTT is 100% local. If you use a cloud provider, you pay that provider directly — WinSTT takes no cut and adds no markup.
Languages
What languages can WinSTT transcribe?
That depends on the model you pick. Pin a language under Settings → Model or leave it on auto-detect.
| Model family | Coverage |
|---|---|
| Multilingual Whisper | ~99 languages |
| NeMo Parakeet / Canary | ~25 languages |
| English-only (e.g. `whisper-tiny.en`) | English |
| GigaAM, T-One | Russian |
What languages is the interface in?
Six: English, Spanish, French, Chinese, Hindi, and Arabic. The interface language is independent of the transcription language.
Hardware & models
Do I need a GPU?
No. WinSTT runs on CPU alone. A GPU is optional and only speeds things up where a platform accelerator is available. On Windows, the DirectML path uses any D3D12-capable GPU (AMD/Intel/NVIDIA) and auto-falls back to CPU when none is present.
Which package should I download?
| Platform | Package | Use when |
|---|---|---|
| macOS Apple Silicon | DMG | You have an M-series Mac. |
| Linux x64 | AppImage / deb / rpm | Pick the package format your distro expects. |
| Windows x64 | WinSTT.exe / portable zip | Use the exe for the normal alpha, or the zip for a self-contained app folder. |
On Windows, DirectML is the right default for almost everyone: it is vendor-agnostic and works whether or not you have a GPU. There is no CUDA Windows installer; CUDA remains a custom/development build path.
How many models are there, and can I switch without restarting?
70+ models across 11 families: Whisper, NeMo Parakeet/Canary/native streaming, Moonshine, Cohere, Granite, Qwen3-ASR, SenseVoice, Dolphin, GigaAM, Vosk/Kaldi/Zipformer, and T-One. All are ONNX and swappable from the UI — most model changes hot-swap live. Switching the backend family or toggling cloud vs local reconnects the engine, which takes a moment.
Can I use my own fine-tuned model?
Yes. Export your fine-tune to ONNX (Optimum / onnx-community layout) and drop it under your user-data models/custom/<slug>/ folder — the model picker has an "Open custom models folder" button that takes you straight there. See Custom Models.
Greyed-out quantizations
WinSTT doesn't quantize models itself — it surfaces whatever the upstream Hugging Face
repo published. If a q4 graph isn't in the repo, that option is disabled.
Data & uninstall
Where does WinSTT store my data?
Settings, the model cache, history database, saved recordings, and the rolling debug.log live under the OS-standard WinSTT app-data folder. Use Open App Data Folder from the app instead of guessing paths. Windows portable mode keeps data in a local Data/ folder next to the executable.
How do I uninstall WinSTT?
Quit the app, then remove the package using the normal flow for your platform: delete the portable app folder, remove the Linux package, or delete the macOS app bundle. To remove everything including downloaded models, history, and settings, also delete the WinSTT app-data folder opened by Open App Data Folder.
Deleting user data is permanent
Removing the WinSTT app-data folder erases your transcription history, saved recordings, and downloaded models — there's no undo. Back it up first if you want to keep any of it.
Related
Troubleshooting
Symptom-to-fix table for the issues people actually hit — blank window, slow transcription, GPU fallback, a stuck mic, missing playback audio, failed downloads, blocked hotkeys, and Listen-mode Bluetooth.
Manual Model Install
Install WinSTT's built-in STT models offline, behind a proxy, or on an air-gapped machine by staging files into the Hugging Face cache.