WinSTTAcknowledgments
The open-source models, runtimes, and tools WinSTT is built on — with links and licenses.
WinSTT stands on a lot of generous open-source work. This page is the friendly summary of who to thank; the full per-file license text lives in THIRD_PARTY_NOTICES.md.
Check the model licenses for your use case
WinSTT is MIT, but the models it can download carry their own terms. Most are Apache 2.0 or MIT, but a few (Cohere weights, NeMo weights) are CC BY-NC / CC BY — fine for personal use, but read the upstream license before shipping anything commercial.
Speech recognition models
Every model is loaded as ONNX through the engine below — WinSTT runs no PyTorch in the transcription path.
| Project | What WinSTT uses | License |
|---|---|---|
| OpenAI Whisper | base, small, medium, large-v3-turbo, and the English-only .en variants | MIT |
| Lite-Whisper | Compressed Whisper variants (-acc, -fast) | Apache 2.0 / MIT |
| NVIDIA NeMo | Parakeet TDT (English + multilingual) and Canary | Apache 2.0 (code); CC BY 4.0 (weights) |
| Moonshine | Efficient ASR for edge devices (Useful Sensors) | MIT |
| Cohere | Open-weight ASR for low-resource languages | CC BY-NC 4.0 (weights) |
| GigaAM | Russian ASR | Apache 2.0 |
| Vosk / Kaldi | Russian small-footprint models | Apache 2.0 |
| T-One | Russian ASR | Apache 2.0 |
| onnx-community / Xenova | The ONNX exports of every model above | — |
Voice activity & wake word
| Project | Role in WinSTT | License |
|---|---|---|
| Silero VAD | Neural voice-activity detection | MIT |
| WebRTC VAD | Fast classical VAD for the pre-check stage | BSD |
| Picovoice Porcupine | Free-tier wake-word engine; built-in keywords under Picovoice's free-tier license | Apache 2.0 (SDK) |
Text-to-speech
| Project | Role in WinSTT | License |
|---|---|---|
| Kokoro-82M | The bundled ONNX TTS voice model | Apache 2.0 |
| kokoro-onnx | Reference implementation that helped validate the ONNX voice model | MIT |
Inference & audio runtime
| Project | Role in WinSTT | License |
|---|---|---|
| ONNX Runtime | The single inference engine WinSTT ships | MIT |
| DirectML | Execution provider that makes the default GPU build vendor-agnostic (AMD / Intel / NVIDIA via DirectX 12) | MIT |
| OpenVINO | The EP for the Intel-tuned build (ARC dGPU / Iris Xe) | Apache 2.0 |
| PortAudio | Cross-platform audio I/O | MIT-style |
| onnx-asr | Reference ASR implementation used by an earlier version of WinSTT and retained for benchmarking | MIT |
Frontend & desktop
| Project | Role in WinSTT | License |
|---|---|---|
| Tauri | Desktop runtime and WebView shell | MIT / Apache 2.0 |
| Vite | Renderer build tool | MIT |
| React | UI library — with the React Compiler Babel plugin that frees the renderer from useMemo / useCallback | MIT |
| Base UI | Accessible unstyled component primitives | MIT |
| Hugeicons | Icon set | Free-license icons |
| Zustand | State stores | MIT |
| use-intl | i18n runtime | MIT |
| uiohook-napi | Global keyboard hook for Push-to-Talk / Toggle | MIT |
Backend & build tooling
| Project | Role in WinSTT | License |
|---|---|---|
| Rust | Native backend language and toolchain | MIT / Apache 2.0 |
| Cargo | Rust package manager and build runner | MIT / Apache 2.0 |
| hf-hub | Hugging Face model download / cache layer | MIT / Apache 2.0 |
| Bun | JavaScript runtime + package manager | MIT |
| ESLint | TypeScript linting, with typescript-eslint | MIT |
| Sentry | Opt-out crash reporting | BSL |
Documentation
- Fumadocs — the framework powering this site. MIT.
Inspiration & adjacent projects
| Project | Why it matters | License |
|---|---|---|
| RealtimeSTT | One of the upstream projects that informed the original WinSTT reference app | MIT |
| Whisper.cpp | The project that proved Whisper could run efficiently on CPU | MIT |
And a personal thanks
Thank you to the community
To everyone who has filed an issue, opened a PR, suggested a model, or asked a sharp question on Discussions. WinSTT is a side project; the community feedback is what keeps it focused. Want to contribute? Start with CONTRIBUTING.md.
Related
Browse the catalog
All 70+ STT models across 11 families, with accuracy, speed, and language coverage.
Compute & quantization
What CPU, DirectML, and OpenVINO mean — and how to read the quantization labels.
Text-to-speech
Read selections aloud with the bundled Kokoro-82M voice model.
FAQ
Privacy, licensing, offline use, and the local-first guarantees.