Slumbr 1.1.1

Slumbr 1.1.1 is a maintenance point-release over 1.1.0 — correctness and robustness fixes from a post-tag code review. No new features; it's a safe drop-in update.

Fixed

• Saved history no longer wiped on restart. With Keep history across restarts enabled, quitting and reopening now merges your in-memory and on-disk transcripts instead of clearing the store, so saved history actually survives a restart.
• Thread-safe history. Every read/write of the history buffer is now lock-guarded, closing a race where a transcript landing mid-rotation could be silently lost.
• Mic-mirror resilience. The virtual-cable mirror reopens after an audio-device error and is guarded against being torn down mid-dictation by a concurrent Settings device change.
• CPU build never hard-crashes on a GPU config. On a machine with an NVIDIA GPU, the CPU installer could end up recommended/configured for the CUDA engine it doesn't bundle and crash on launch. It now transparently falls back to the built-in Moonshine CPU engine (and points you to the NVIDIA build for GPU acceleration).
• Install from paths with spaces. The source/editable install command is built as a proper argument list, so installing from a folder whose path contains spaces no longer fails.
• Atomic state + safer shutdown. The transcriber waits for its worker before closing (avoids a native use-after-free), and the history store quarantines a corrupt database and survives Windows file locks.
• CPU exe now identifies as "Slumbr 1.1.1" in Explorer/taskbar (the version resource is embedded, matching the NVIDIA build).
• Runaway-repetition collapse thresholds pinned with regression tests.

Downloads

• slumbr-setup-cpu.exe — runs on any Windows 10/11 PC (CPU, built-in Moonshine engine). Fully offline on first run.
• slumbr-setup-nvidia.exe — for NVIDIA GPUs (faster-whisper on CUDA, large-v3-turbo / large-v3). Downloads the GPU Whisper model on first run.

A note on SmartScreen

These installers are unsigned (code-signing is pending an organization certificate). Windows SmartScreen may show a "Windows protected your PC" prompt — click More info → Run anyway. The source is fully public if you'd rather build it yourself.

See the full CHANGELOG.

Slumbr 1.1.0

Slumbr 1.1.0 — a post-launch tune-up driven by community feedback: smarter history, an opt-in to keep it across restarts, an accuracy tier for strong GPUs, and real CI behind the scenes.

Highlights

• History is now a rolling window of your latest 200. Past the cap it drops the oldest entry instead of wiping the whole list — your recent dictations are always there.
• Opt-in: keep history across restarts. Settings → History → Keep history across restarts saves transcripts to a local file so they survive a reboot. Off by default — the privacy-first, in-memory behavior stays the default, and turning it back off deletes the file.
• Full large-v3 accuracy tier on high-VRAM NVIDIA GPUs (≥10 GB). The most accurate Whisper model, offered as the top Engine pick where it still runs in ~1 s. Mid-range cards keep the fast large-v3-turbo default.

Under the hood

• Loosened dependency pins so installs-from-source don't fight the resolver on routine updates.
• Added a Windows CI workflow that runs the test suite on Python 3.10 / 3.11 / 3.12 on every change.
• Docs: a "Why Slumbr?" comparison, a vocabulary-hint explainer, transcription-failure behavior, and a clearer note that the optional reverse-PTT (VB-Cable) is the only part that needs admin.

Install

Grab the build for your hardware below. Not sure? Start with the CPU build — it runs everywhere.

Slumbr is open-source and not code-signed yet, so Windows SmartScreen may flag it on first run. Click More info → Run anyway. (Signing is on the way.)

Thanks to ThatOneGuy for the detailed review that shaped this release.

Slumbr 1.0.0

Slumbr 1.0.0 — local, offline voice-to-text for Windows

Tap Caps Lock, speak, tap again — your words type into whatever window is focused. Fully on-device: no accounts, no cloud, no telemetry.

Download

Installer	For	Notes
slumbr-setup-cpu.exe	Any PC	Runs on every x64 Windows machine (NVIDIA / AMD / Intel / no GPU). CPU-speed. Bundles its models — works offline the moment it's installed. Start here if unsure.
slumbr-setup-nvidia.exe	NVIDIA GPU	GPU-accelerated (CUDA). Bundles the CUDA runtime + the live-preview models; downloads the GPU transcription model (~1.5 GB) once on first launch.

"Windows protected your PC"? Slumbr is open-source and not code-signed yet, so SmartScreen warns on first run — click More info → Run anyway. Or install from source (see the README) for AMD/Intel GPU acceleration.

If a build doesn't match your GPU, Slumbr runs on its built-in CPU engine and points you at the right build — it never leaves you stuck.

Highlights

• One-click installers that work on any PC, plus a from-source path that auto-detects your hardware (NVIDIA CUDA · AMD/Intel DirectML · CPU).
• Offline first run — the CPU build ships its speech models inside the installer, so there's nothing to download before you can dictate. (The NVIDIA build downloads its GPU model once on first launch.)
• Live partials while you speak (Moonshine + Silero VAD + online punctuation) and an accurate final transcribe from your chosen backend.
• Tap-to-toggle hotkey (rebindable), auto-paste at the cursor, and copy-from-History (in-memory only — nothing about your dictations is written to disk).
• Mute other apps while dictating (VB-Cable routing — Discord/Zoom/Teams/OBS/browser calls).
• Resilient startup: if a GPU backend can't load on your machine, Slumbr falls back to the CPU engine instead of failing.
• 100% offline at runtime — after the one-time setup, nothing leaves your machine.

Requirements

Windows 10/11 (64-bit). GPU optional — the CPU build runs anywhere. Full details in the README.

Full changelog: see CHANGELOG.md.