The full version for macOS, with auto paste and every feature.
Download now v0.7.3Also on the App Store (limited version: no auto paste and fewer features). Learn why.
whisper.cpp now runs inside the app, built from our own pinned source and updated to v1.9.2, with the Metal shaders precompiled. The engine warms up when xcribe launches, so even your first dictation no longer waits on model loading. Dictations are faster end to end.
A new toggle under the model picker, for whisper models. On, the model stays resident and dictations are instant. Off, xcribe loads the model per dictation and frees the RAM in between. The toggle shows each model's real size, so you know what you're spending. It ships off by default.
Parakeet TDT v3 can now transcribe in a language you choose. Its old auto-detection mixed languages mid-sentence, so it now runs with an explicit language, English by default. For automatic language detection, whisper is the one to use.
OpenAI and custom-compatible providers get a Request Parameters section in the edit dialog: temperature, penalties, token limits, and more. And when a provider rejects something, you now see its real error message instead of silent retries.
xcribe records straight from your selected input device and never touches the macOS system default anymore. Before, it fought with Sound Settings and you had to quit the app to take control back.
New users start on whisper small, the transcription error dialog now comes to the front, and language pickers no longer jump around while you browse them.
xcribe is a labor of love, built entirely in my free time. If you find it useful and want to support continued development, new features, and improvements, consider buying me a coffee. Every bit helps keep this project going!
☕ Buy me a coffee