murmur 1.0.0

murmur 1.0.0

Hold Ctrl+Win, talk, release. Local faster-whisper types what you said. No AI rewriting on the
dictation path, ever.

Since 0.9.0

• Promptify: turn a dictation into a written prompt with the engine you already pay for
  (Claude, Codex, Gemini, OpenRouter or a custom URL), never on the dictation path.
• An Obsidian vault bridge that gives Promptify your own notes as context, opt-in and disclosed.
• Trigger key grammar: bind any key - a headset's Play/Pause, a media key, F13. A single tap
  still does the key's own job, double-tap starts and stops, triple-tap opens the window, a real
  hold records while held.
• Three waveform styles in Settings > Appearance: glass ribbon, glass equaliser, fast liquid
  glass, plus a theme that follows Windows.
• A Fluent-2 window rebuilt end to end, with screenshots and a first-five-minutes README.
• Whisper repetition loops after a filler and a long pause are detected by compression ratio and
  re-decoded once, decode-side only.

Under the hood

• Install 299 -> 175 MB; view switches 4-6x faster; the overlay renders at device pixels
  (36-49 -> 4.9-6.7 ms a frame at 200 %) and runs at 30 fps.
• A long take no longer holds audio it has already transcribed: a 30 minute take's peak memory
  244 MB -> 1.2 MB. Startup to the tray icon 2.0 -> 1.0 s.
• A second launch opens the running window instead of starting a second copy that pasted
  everything twice; every unhandled exception now reaches murmur.log; config values are checked
  by type and saved atomically.

Seven test suites, no framework. Windows only.

murmur 0.9.0

Everything since 0.2.0, in one release. Dictation is unchanged in principle: hold Ctrl+Win, talk, release, the words are typed verbatim - local Whisper, no cloud, no rewriting.

New: Promptify

A rambling take → a prompt for Claude, inside the app. Pick a dictation in the new Promptify section (or press Promptify / Ctrl+D on a History entry): the goal first, in your words, your terms and decisions kept verbatim, musings kept as open points, and at most three questions about the things only you can answer. Answer by clicking a chip, typing, or holding the chord and talking into the field; Update prompt folds the answers in; Copy prompt, paste into Claude Code (or claude.ai). The draft stays on the entry. Only the dictation you pick is sent, only when you press the button.

The engine that writes the prompt signs in on its own, in your browser - Promptify › Engines › Connect: Claude Code (your claude.ai subscription), Codex (your ChatGPT plan), Gemini CLI (your Google account) or OpenRouter (one sign-in that reaches Claude, GPT and Gemini models, free ones included). murmur keeps no keys: the three CLIs hold their own logins and murmur only runs the binary you signed into; the OpenRouter key its connect flow hands back is the one thing stored, in %APPDATA%\murmur.

The window

Rebuilt to a written design method: History · Promptify · Settings in a sidebar, a two-pane Promptify view with one primary action, light and dark from the Windows theme, a white ground with the brand green as the accent. The identity is the Ripple m - the tray icon tinted like the bar (accent while the mic is open, the transcribing colour while it types), the title bar and installer icons, the sidebar lockup.

Dictation, since 0.2.0

• Streaming: transcription runs while you talk, cut only where Whisper ends a sentence; text lands ~9 s after release on a 61 s take instead of 20-26 s.
• The overlay is a Win32 layered window (DPI-aware) with a waveform or a breathing light, your colours, opacity and haze; one clock-based fade in and out.
• A bindable trigger key (headset button, media key, F13) toggles recording; double-tap Ctrl+Win for persistent mode.
• Microphone kept by name, never by index (Windows re-enumerates); the log names the mic it opened.
• Performance: priority boost while transcribing, offline-first model load, deterministic decoding; startup 16 s → ~4 s warm.
• History window with retention, copy, delete with undo; the log records timings and word counts, never the words.

Install

Run murmur-setup.exe (per-user, no admin). First run downloads the Whisper model once (~480 MB); after that it works with Wi-Fi off. Windows 10/11.

murmur 0.2.0

Hold Ctrl+Win to dictate; double-tap for persistent mode; optional wired-headset button. Local faster-whisper, no cloud, no AI rewriting. Per-user installer, no admin. Unsigned: SmartScreen will warn - More info > Run anyway.