kuular.
Get Kuular
Voice-to-text, reinvented. Kuular runs GPU-accelerated AI right on your PC
and turns what you say into clear, ready-to-use text — without the usual struggle.
On-device AI. No account required. Your words stay yours.
FROM “I SHOULD WRITE THAT DOWN”
For the first draft, the next big idea, and everything you don't want to forget.
Record from Studio or use your own global shortcut. Optional live previews let you follow your words as you speak.
Hold to record, or double-tap to start. Your shortcut, your choice.
Search your local transcript library by text, status or date. Open the full transcript, copy it, or replay the original audio.
Discover compatible Hugging Face models. Choose your balance of model size, language support and hardware.
Load GGML models on CPU or Vulkan, or use optional PyTorch with NVIDIA CUDA. Named device selection puts you in control of where inference runs.
Sherpa-ONNX runs on CPU. GPU support depends on the runtime and driver.
NO MOCKUPS. THE REAL WORKSPACE.
Actual application screenshots with demonstration data. No personal recordings shown.
SOME THINGS SHOULD STAY WITH YOU
Ideas can be messy. Personal. Not quite ready for the world. Kuular transcribes locally, so the audio you record doesn't need to leave your computer.
Internet is needed to download the app and models. Optional model discovery contacts Hugging Face; it does not upload your recordings.
A SMALL RITUAL. A BIG DIFFERENCE.
Install Kuular on Windows and download a compatible local speech model.
Pick your microphone and press record. Or stay in your flow with a global shortcut.
Stop recording to finalize your transcript. Review, copy and revisit it in your library.
A FEW THINGS WORTH KNOWING
Thoughtful answers, before you press record.
Kuular's speech recognition runs locally. Recordings, transcripts and timestamps are saved in your local database. Downloading models and browsing Hugging Face requires internet access, but does not send your audio there.
The initial release is for Windows x64. The self-contained installer includes the .NET runtime. Whisper.cpp also requires the Microsoft Visual C++ 2015–2022 x64 redistributable, installed separately; that Microsoft installer may require administrator approval. You'll need a microphone, disk space for the app and a compatible model, and enough memory for that model. A GPU is optional. Model memory requirements vary; weight sizes are not total runtime memory requirements.
No. GGML and Sherpa-ONNX support CPU transcription without Python. GGML can use a compatible Vulkan GPU. Transformers is an optional advanced runtime requiring a separate Python environment; NVIDIA CUDA needs compatible hardware and drivers. That environment and speech models are not bundled with the installer.
No. Live previews are revisable text from a rolling recent-audio window. After you stop, Kuular transcribes the full recording and saves the final result. It isn't a word-by-word streaming decoder.
Speech recognition can make mistakes. Accuracy depends on your model, language, microphone and environment. Review important transcripts, especially names, numbers and sensitive information. Your saved original audio lets you check what was said.
This is an early Windows build. The installer is not code-signed, so Windows SmartScreen may display an unknown-publisher warning. Check the published SHA-256 checksum and only install a download you trust. macOS, speaker diarization and system-audio capture are not included.
GIVE YOUR NEXT IDEA A VOICE
Your private transcription workspace.
Ready for the way you think.