0.6.2

Two new models

• Granite Speech 5.0 470M — IBM's new English speech model, in Models → Speech. 546 MB, Apache-2.0, and fast: about 44× realtime on CPU. IBM ships it as PyTorch only, so the ONNX conversion is Murmur's own, verified against IBM's reference from 1.3 to 18 seconds of audio and hosted on its own release with every file pinned by SHA-256.
• Granite 4.2 3B — IBM's latest polishing model, in Models → Polishing. First-party Apache-2.0 build.

Both are downloads, not replacements: your current models stay selected until you pick something else.

Fixes

• Speech models no longer invent words during silence. Handed audio with nobody speaking, every ONNX model here would produce something — usually "thank you" or "okay" — and it landed in meeting transcripts. Murmur now checks whether anyone actually spoke before asking the model, using the same voice detection the rest of the app uses. Quiet speech is unaffected: it is still recognised at a twentieth of normal volume, buried in noise, and in utterances as short as a single word.
• Switching desktops no longer drags you back. Swiping to another Space could snap you to the one Murmur was on, with Murmur in front and no key ever pressed — the app was activating itself as a side effect of building its own window. It cannot any more.

---

Speech and polishing servers are bundled, so a fresh install can transcribe once you download a model from the Hub.

macOS is signed with a Developer ID and notarised, so it opens normally.

Windows is not signed. SmartScreen will warn: More info → Run anyway. Murmur asks to watch your keyboard and type into other applications, so only accept that if you trust this build — building from source sidesteps it.

Linux needs an X11 session (not Wayland) and glibc 2.39+ — Ubuntu 24.04+, Debian 13+, Fedora 40+.