We can try out Vosk in vocalinux where that does real-time dictation but might not be as accurate as Whisper models.

Been building in public it for last eighteen months almost. Bit by bit, with the community.

We have pushed a lot of Ibus/Wayland/X11 fixes in last couple of months. Would you mind trying it again on Arch? It should be able to inject it properly now.

Works fine next to a notes app. Dictation goes into whatever has focus.

Glad the mic gain fixed it. 8 GB is tight for the bigger models, so sticking to tiny/base or quantized whisper.cpp is the sane path there. I’m also adding support for Parakeet models soon. That works incredibly well on CPU only machines with low RAM.

Yeah, pipe-to-bash is a bad habit even when the script is ours. Prefer git clone / AppImage / read install.sh first; the /tmp step is so you can open it before it runs, not so you skip reading it.

Fair if whisper.cpp tiny/base feels dated. Default is still on-device whisper.cpp with Vulkan, larger models and VOSK are in Settings, and we are adding Parakeet. Code is AGPL on GitHub.

Also, I am a backend person first, so the old site taste was on me. I have been working on that. Would love you to hit vocalinux.com again and tell me what still sucks. :)

Thanks for the Vocalinux shoutout. I’m the creator and maintainer.

We also have an Android build now, VocaPhone: on-device speech to text by default, no LLM in the loop. Public beta APKs are on GitHub if you (or OP) want to try it: https://github.com/VocaHQ/vocaphone/releases/tag/android/v0.1.6

Site: https://vocaphone.vocahq.com/

Working on getting it onto the Play Store.