What's new

The Last transcript card now shows what really happened to your clipboard, discarded or timed-out recordings keep their stop reason, and the paste, focus and clipboard code moved into a platform layer as groundwork for Linux and macOS builds.

New

• Record from the microphone on Linux (#63)
• Paste, type out and transform selected text in X11 apps on Linux (#64)
• Hotkeys that work on Linux for new installs (#66)

Fixed

• A discarded or timed-out recording keeps its stop reason (#65)
• Show what really happened to the clipboard on the Last transcript card (#69)

A reorganized Scribe: settings are grouped, Ctrl+K finds any setting, and Home shows your stats. Scribe's windows stay responsive while it transcribes, and settings changed in one window are no longer undone by another.

New

• Grouped sidebar with Ctrl+K settings search, stats on Home, History and Notes as tabs (#57)
• Show release notes before updating, once after an update, and in About (#55)
• Copy a diagnostics report from About for bug reports (#58)

Fixed

• Keep every Scribe window responsive while a dictation transcribes (#53)
• Settings saved in one Scribe window are no longer undone by another (#54)
• Tidier Home cards and plainer wording on Hotkeys, Local AI and Models (#61)

Dictations now paste only into the window you started in, and stay out of Windows clipboard history. New in Settings: choose what happens after dictation (insert, copy only or save only) and how text goes in, including typing it out for apps that block paste. Web pages can no longer reach Scribe's local speech server.

Fixed

• Align GitHub backup setup guidance
• Clear the Dependabot alerts (#40)
• Transform Selection only rewrites what you selected, and keeps your clipboard (#48)
• Clear leftover recordings and old data, and keep settings when one is unreadable (#47)
• Hotkey conflicts on the Dashboard, expired paste buffer, Stats clipping, MSI smoke hang (#50)

New

• GitHub backup signs in through Scribe's own GitHub App, with access to only the repo you choose.
• The Sync screen shows whether your backup is working, with a history of each automatic backup.
• Restore transcripts from a Scribe JSON export. Scribe checks the whole file and shows what it will add before it changes anything.

Fixed

• Cancelling GitHub sign-in partway through now stops it cleanly.
• Installing an update only closes Scribe's own processes, never anything else.
• Backup setup matches the permissions the GitHub App asks for.

Safer

• Every speech model download is checked against its known SHA-256 hash.
• CSV exports can't carry spreadsheet formulas from a transcript into Excel or Sheets.
• The local dictation-state server's access file is readable only by you, and the server limits how much it accepts.
• Vulnerable Rust and npm dependencies were updated.
• Each installer is test-installed on Windows and checked against its published checksums before a release goes out.

New

• Other apps can see when you're dictating: Scribe shares its live state (idle, listening, busy) through a local, token-protected server, so a tool can hold its own sounds or messages while you talk.
• A small scribe command-line client and an MCP server read that state, for scripts and AI assistants.
• Scribe also writes its current state to a status file on disk (0.6.1).

Fixed

• The floating pill can no longer get stuck off-screen after a minimized window or an unplugged monitor.

GPU acceleration, and a simpler/faster local-LLM story.

GPU acceleration (Vulkan)

• Whisper now runs on your GPU. The bundled whisper.cpp is built from source with Vulkan, so the large/most-accurate models run dramatically faster - on a Radeon RX 7800 XT, large-v3-turbo went from ~12 s to ~0.75 s on an 11 s clip (~16× end-to-end, ~88× on the encode), with identical output. Works on any modern NVIDIA/AMD/Intel GPU; falls back to CPU automatically when there's no usable GPU, and a GPU failure retries on CPU rather than losing the dictation.
• On by default (Auto) - existing installs start using the GPU after updating; turn it off (or pin a specific GPU on multi-GPU machines) in Audio → GPU acceleration, which also shows the detected device. The installer download grows only ~7 MB despite the new Vulkan backend (the shaders compress well).
• Full-precision large-v3-turbo added to the model catalog alongside the quantized one (max accuracy, GPU-recommended), plus an English / Multilingual filter in the model browser.

Text quality & speed

• Silent recordings no longer paste invented text. A record toggle on/off with nothing said now returns empty instead of a Whisper silence-hallucination - the model is skipped entirely when no speech is detected.

Removed

• AI dictation cleanup (local LLM) is gone. It duplicated the deterministic filler suppression and added a per-dictation LLM round-trip (a real slowdown) for little benefit. Use Filler suppression (instant, on-device) instead. The local LLM still powers Notes analysis and Selection transform - both on-demand, never in the dictation path.