Your voice, on your machine.
Offline dictation and read-aloud for the desktop.
Built with Electron, React, TypeScript, and sherpa-onnx.
Downloads · Contribute · Report an issue
Watch the full recording with sound (about 4 minutes, recorded on Windows from a real run; History text is blurred).
- Dictate with a keyboard shortcut and keep a searchable transcript history.
- Transcribe what your computer plays, such as a meeting or a video, as it is heard. Grey provisional text keeps up with the speech, and a small floating window shows the transcript without opening Inaudio.
- Pick the source: system audio, system audio with your microphone, or one audio device such as a virtual cable.
- Clean up noise on the device: a voice detector skips stretches without speech, and an optional speech cleanup model filters background noise.
- Read text or your clipboard aloud with on-device speech synthesis.
- Manage speech models, microphone settings, and playback in one desktop app.
Download a model once, then run inference locally. Model downloads need an internet connection. Desktop packages target Windows, macOS, and Linux; permissions and text insertion vary by platform. Releases are distributed through electron-builder update manifests; CI artifacts are currently unsigned and development builds do not contact the update feed.
Speech-to-text includes NVIDIA Nemotron ASR Streaming, an English cache-aware streaming model with a 560 ms latency/accuracy profile. It runs through sherpa-onnx on the CPU after its verified model archive is downloaded; the model remains subject to the NVIDIA Open Model License.
Use Node.js 24 LTS and npm. Native packaging prerequisites are in CONTRIBUTING.md.
git clone https://github.com/shikki841/inaudio.git
cd inaudio
npm ci
npm startnpm run check # TypeScript + ESLint
npm run package:release # Build the electron-builder package for your current OSPick a model during setup and grant microphone access when prompted.
Small fixes, platform testing, and clear bug reports help. Read the contribution guide, community guidelines, and release process. Report vulnerabilities through the security policy.
MIT. Downloaded models and dependencies retain their own licenses.