Lataa murmure 2.0.3

Murmure 2.0.3
Murmure 2.0 is the biggest update so far: a new interface on every page, a more accurate transcription model that uses less memory, and Command Mode and the local API as their own extensions.
A new interface
Every page was redesigned, with new fields, tabs and buttons. Settings now comes before Extensions, the getting started checklist moved to the sidebar, and each extension page has its own color, with its on/off switch next to the title. The interface also shows right away at startup, while the model loads in the background.
Parakeet ultra
Murmure now uses Parakeet ultra, a fine-tune of Parakeet by Moondream, in the int8 export by @thiswillbeyourgithub. It makes fewer mistakes and uses about 380 MB less RAM.
Command Mode and Prompt Mode
LLM Connect is now called Prompt Mode, and Command Mode is a separate extension with its own model. With nothing selected, a command answers your question instead of repeating it. In Prompt Mode, the server options are now named Ollama and OpenAI-compatible, so LM Studio, Magnitude or llama.cpp work through the OpenAI-compatible option.
The local API
The local API is no longer experimental. It has its own extension page, new endpoints that apply a Prompt Mode prompt to the transcription, and a Try it button to test each endpoint with a WAV file.
Also new
- The result can stay on screen after a dictation, with a copy button.
- A new Automatic insert switch, separate from the insertion method. When it is off, Murmure never types the text by itself.
- A --quit flag to close Murmure from the command line.
Fixes
- Voice Mode triggers less often by mistake. Short words such as "ok" must now match exactly.
- Hesitation sound removal is optional and off by default, as it also removed real words such as the German "um".
- Transcriptions no longer contain a stray <unk> token.
- On Windows, the title bar, scrollbars and native controls are now dark.
- On macOS, clicking the Dock icon reopens the window, and the overlay shows over full-screen apps.
- On Linux, the volume of audio started during a dictation, such as a new browser tab, is restored at the end.
- Downloaded updates keep their real file extension.
- The tray menu is translated.
- Security hardening: the Smart Mic pairing token moved to the URL hash, the local API rejects requests from a foreign origin, recording files are owner-only, the .deb is checked before install, the Ollama URL is validated, and dependencies got their security updates.
🙏 Thanks
- Beta testers: @laurentkempe, @Choupinounours, @ldmpub, @Martin511, @HachetAmaury, @anaismarongiu
- Contributors: @xop01, @linkermaxx, @thiswillbeyourgithub
- Donors: Migraine, Ubertartine, Chase, Gizmo38, Aparrot, Prometee, dmxplx, DrROBINN
See full CHANGELOG.