← Back to site

Release history

Changelog

Every version below is a tag that exists, dated the day it was cut. Nothing here is planned work — if it is on this page, it shipped.

Releases
24
Latest
v0.6.3
Shipped
September 17, 2026
Every release on GitHub

v0.6.3

Current

View tag ↗

Models silently running on the wrong GPU, and two ways a dictation could vanish on its way to the window you were typing into.

Fixed

  • GPU acceleration quietly landing on the wrong GPU is now visible. A driver update can leave the dedicated card unregistered with Vulkan while the integrated one stays visible — the load still reports success, but the model is in system RAM and several times slower. OmniVox logs which device it used, and says so once when the only one is integrated.
  • The overlay pill can no longer take focus from the window you are dictating into. Clicking it made the pill the foreground window, and OmniVox refuses to paste anywhere but the window you started in — so those dictations disappeared without a word. A refused paste now shows on the pill itself.
  • Hotkey chords that start with Alt no longer lose every other dictation. The app you were typing into saw a bare Alt tap, which in Chromium and Electron windows moves focus to the menu bar, and the text landed there instead. OmniVox now covers that tap with an unassigned keystroke while the modifier is still held.

Hardened

  • rustls moved to 0.23.45 for RUSTSEC-2026-0285. TLS is only used for model downloads and the optional AI meeting notes, so nothing is asked of you.

v0.6.2

View tag ↗

Fixed

  • The v0.6.1 installer was the wrong file. Packaging picked its installer with a wildcard and a cached build from an earlier run won, so the still-broken bundle went out under the v0.6.1 name. It now pins the exact filename for the version being released, and refuses to publish an installer whose archive is missing the runtime libraries.

v0.6.1

View tag ↗

Fixed

  • Fresh installs of v0.6.0 could not start. The installer put the speech-engine runtime libraries in a folder of their own instead of beside the executable, so any machine without a previous install failed at launch with a missing-DLL error. Installing over v0.6.0 fixes the layout; nothing else changed.

v0.6.0

View tag ↗

Meeting Mode, a second speech engine, and a cleanup pass that turns a dictation into written text before anything downstream sees it.

New

  • Meeting Mode. OmniVox captures your microphone and the system audio together — Meet, Zoom, Teams, Discord, whatever is playing — in recoverable 20-second chunks that Whisper transcribes on your own machine and deletes the moment their text lands.
  • A widget docks to the edge of the screen while a meeting runs, doubling as the recording indicator and the way into your notes. The notes themselves open in a drawer you can float over any app: rough notes beside a searchable transcript, with markers, speaker labels and tags.
  • OmniVox notices when you are in a call. A background check watches which apps hold the microphone and corroborates it against your open windows — every rule wants two pieces of evidence — then offers to start recording. Nothing records until you say so.
  • Optional AI meeting notes, through an OpenRouter key you supply yourself. Summaries, decisions and action items, each carrying a timestamp back to the line of transcript it came from. Inactive until you enter a key, which is kept in Windows Credential Manager, and audio is never uploaded.
  • Parakeet TDT, a second speech engine beside Whisper. Download and activate it the way you already do and the whole engine swaps behind the pill. It is English-only and CPU-only, and vocabulary bias and Screen Context do not reach it — those condition Whisper's prompt. Whisper stays the default.
  • Cleanup Mode. A small local model rewrites a dictation as written text — fillers dropped, spoken self-corrections resolved, punctuation, casing, numbers and dates applied — before Structured Mode, list formatting, your output and your history ever see it. Off by default, English, CPU.
  • Distil Large V3.5 joins the speech catalog. Distil Large V3 is still what the app recommends for high-tier hardware.
  • Choose how long history is kept: until you delete it, or 7, 30 or 90 days. Older entries are removed in the background when you change it. New installs keep 30 days; existing installs keep whatever they had.

Improved

  • The largest visual pass since the app shipped. Ten shared primitives replace controls that had drifted — native checkboxes and dropdowns that ignored every theme token, five separate progress bars, confirm dialogs whose only button said OK.
  • Pages arrive as skeletons shaped like their real content instead of one centred spinner. The dictation home is optically centred, with the record button holding one position whether or not a result is on screen, and the sidebar is grouped rather than nine flat items.
  • Fonts ship with the app. Geist and Geist Mono were being fetched from Google on every launch, from three separate windows. They are bundled now, so the app draws its own text with no network at all.
  • The light theme is legible. The dark colour ramp was being used verbatim on the light canvas, which left command-mode badges all but invisible; the recording red, the command blue, the semantic role colours and the success green are remapped for it now. Every shadow in the app was silently dead in light mode as well.
  • OmniVox holds far less memory when you walk away. Structured Mode's model weights are released after 45 minutes idle — roughly a gigabyte that used to stay resident for the whole session — and Whisper's retained decode state after five. The next dictation after a long gap pays one cold load.
  • “8 seconds” now means 8 seconds. With Cleanup and Structured Mode both on, the two stages each claimed a full timeout, so an 8-second setting could hold a dictation for 16. They split one budget now.

Hardened

  • Model downloads are verified. Every artifact is pinned to an upstream revision with an expected size and SHA-256, streamed to a temporary file and checked before it is published under its real name, so a truncated or tampered download cannot pass itself off as an installed model. Models from older releases are verified once on upgrade.
  • Every release publishes a checksum manifest, and the install script refuses to run an installer it has not verified against it — pinned asset names, a github.com download URL, the size GitHub reports, then the published digest. Releases are still not code-signed.
  • Each window gets only the permissions it needs. The single capability that handed event-emit and full window control to the main window, the overlay and the Scratchpad alike is replaced by six per-window manifests — the overlay can only listen, and only the windows you actually drag hold the dragging permissions.

v0.5.0

View tag ↗

A place to put passing thoughts, a Command Mode you can drive from the keyboard, and Structured Mode in three shapes instead of one.

New

  • The Scratchpad — an always-on-top pad you can dictate into from any window, as separate cards or one growing note. It remembers its size and position, and its contents survive a restart.
  • Command Mode answers to the keyboard: Enter confirms, Esc dismisses, and anything that sends a message shows an editable confirm seeded with what was heard — so a mishearing gets fixed before it goes out.
  • Say “stop” to abort a command mid-run, or “undo that” to reverse the last one — reopening a closed app, restoring a minimised window, undoing a paste.
  • Structured Mode profiles. Pick agent-prompt, email or notes-outline per context mode; each has its own extraction shape and formatting.
  • Dictate real lists by voice — “bullet point”, “number item”, “end list” — in plain dictation, with no model involved.
  • Filler removal is now a setting, so you can transcribe verbatim when you are dictating a quote.

Improved

  • A visual pass across the whole app — pages, settings and the overlay — onto one graphite surface with a single amber accent.
  • The overlay pill stops disappearing over a long session. It now re-asserts itself when Windows demotes it, and pulls itself back onto a connected monitor.
  • Long dictations no longer lose content silently in Structured Mode: the input cap rose, and the panel says how much did not fit rather than dropping it without a word.
  • Structured Mode is ready when you are — the model reloads in the background as you start speaking instead of on the next dictation's critical path.

Hardened

  • Every action with a side effect is now tied to a specific window and the process that owns it, and re-checks both before it fires. A window that has been recycled or refocused fails closed instead of acting on the wrong app.
  • Opening a site by voice requires the address to have been in what you actually said; anything else stops at the confirm pill instead of opening blind.

v0.4.0

View tag ↗

Voice Command Mode — speak an instruction and OmniVox carries it out instead of typing it.

New

  • Launch or switch to an app, run a search, open a site, and fire the common keyboard, media and window shortcuts. One sentence can chain steps.
  • Put text into whatever app is focused: “type …” leaves it for you to review, while “tell Claude to …” pastes and presses Enter. Anything that sends goes through a confirm first, so a message is never fired blind.
  • A Commands page listing every command with its own toggle, so you can see what Command Mode understands and turn off what you don't want.

v0.3.1

View tag ↗

Vocabulary entries stopped being a hint and became a correction.

New

  • Adding a word used to only bias Whisper toward it, and “Claude” still came out as “cloud”. A phonetic pass now runs after every transcription and rewrites anything that sounds like one of your entries to the entry exactly as you spelled it — casing included.

Improved

  • Structured Mode got much faster: the model keeps its system prompt cached between dictations instead of reprocessing it every time.
  • Cleaner audio into Whisper. Microphone capture now goes through a proper anti-aliasing filter before downsampling, so hiss and keyboard noise stop folding back into the speech band.
  • Stopping a recording no longer waits on an in-flight live preview before starting the real transcription.

Fixed

  • Long dictations no longer break Structured Mode partway through generation.
  • Structured Mode works again on installs from before v0.2.10, whose model files the catalog rename had orphaned.
  • No more invisible fallback to the CPU. If a GPU load fails, the overlay says so instead of quietly running several times slower.

v0.3.0

View tag ↗

Settings, redesigned.

Improved

  • Grouped cards with scannable rows instead of one oversized card per option, in two columns — roughly half the scrolling.
  • The controls people actually reach for lead: the push-to-talk shortcut and GPU acceleration moved to the top.
  • The record button stopped shifting when a recording starts, and page headers lost their dead space across History, Dictionary, Modes, Notes, Models and Analytics.

v0.2.11

View tag ↗

The Analytics page.

New

  • Lifetime words, day streaks, speaking pace and a contribution heatmap, plus peak hours and a 30-day trend — all computed on your machine from history that never leaves it.

Fixed

  • Amber accents are legible in light mode. The brand scale was tuned for dark surfaces and washed out on the light theme.

v0.2.10

View tag ↗

Fixed

  • Structured Mode model downloads work again — the catalog had been pointing at files that do not exist upstream, so every download failed instantly.
  • A failed download now keeps its error on the row instead of silently resetting.

v0.2.9

View tag ↗

Fixed

  • Screen Context now skips agent and browser-host processes that could hang when read, which had been freezing dictation into those apps.

v0.2.8

View tag ↗

New

  • A Reset Pill item in the tray, for every way the overlay could go missing — ghost mode left on, a monitor unplugged with the pill parked on it, or a fullscreen app stealing the z-order.

v0.2.7

View tag ↗

Fixed

  • Type output no longer clobbers your clipboard. Something copied before dictating is now captured and put back afterwards.

Improved

  • Settings spells out what each output mode does, so the difference between Type and Both is visible without reading the source.

v0.2.6

View tag ↗

New

  • Screen Context, and a round of stabilisation on the dictation pipeline.

Improved

  • Refined design tokens, plus Rust health checks and a dependency audit running in CI.

v0.2.5

View tag ↗

New

  • A Raw button in the Structured panel — an escape hatch that pastes exactly what you said when the model misreads it.
  • The Models page split into Speech Recognition and LLM Structuring tabs, with the Structured Mode controls moved in beside the models they configure.

Improved

  • A polish pass on the pill's animation, chasing down three separate one-frame flickers on expand and collapse.

v0.2.4

View tag ↗

Fixed

  • The Structured panel could fail to appear after the model finished, leaving nothing to paste. It now mounts reliably.
  • More “Voxify” aliases, because Whisper swallows the middle vowel when the word is said quickly.

v0.2.3

View tag ↗

Fixed

  • Right-clicking the pill no longer flashes the menu in the wrong place for a frame.
  • The “dictation too short” banner no longer clips the context menu open behind it.

v0.2.2

View tag ↗

Fixed

  • The Windows release build, again — the build was picking up the wrong linker from the shell's path.

v0.2.1

View tag ↗

Fixed

  • The Windows release build, which had been racing on one of the native dependencies. Also meaningfully faster.

v0.2.0

View tag ↗

Structured Mode — dictation stopped being a transcript and started being a filled-in prompt.

New

  • A local model turns raw dictation into a slot-filled Markdown prompt. Decoding is held to shape by a grammar, so it cannot drift into prose, and it falls back to plain text on timeout rather than failing.
  • It reads intent: an implementation ask gets goal, context, constraints and files; an exploration gets questions; advice gets options. The model fills the slots that fit instead of padding every field.
  • The preview panel that flows out of the pill, with the transcript, Edit, Paste, Copy and Dismiss, a raw-transcript drawer, and a mic for dictating additions straight into it.
  • The Ley Line toggle on the pill's menu, and the “Voxify” gate — say the word at the end and the dictation gets structured; leave it out and it stays plain.
  • A download manager for the Qwen3 structuring models, tracking what is on disk.
  • Fabrication defences: the model is held to what you actually said, so it cannot invent a filename, a constraint or a feature you never mentioned.

Improved

  • A redesign of the pill overlay — warm charcoal surfaces, rim light and grain — and a transition that reads as the panel flowing out of the pill rather than appearing beside it.

v0.1.9

View tag ↗

Fixed

  • Less eager list formatting, and terminal auto-send stopped firing when it shouldn't.

v0.1.8

View tag ↗

New

  • The vocabulary system, audio ducking controls, and prompts that can be swapped without a restart.

v0.1.7

View tag ↗

New

  • A “send” command you can say at the end of a dictation to submit, with its own toggle — so you decide when a message goes out.
  • Creating a context mode now drops you straight into editing it, with dictionary, snippets and app bindings ready.

Fixed

  • An out-of-memory crash during dictation on 16 GB laptops.

v0.1.6

View tag ↗

New

  • The first voice commands — new line, new paragraph, delete last word.

Improved

  • Clipboard paste output made reliable, plus a fix for chunked typing and a round of performance work.

v0.1.0

Six earlier tags — v0.1.0 through v0.1.5 — were cut in the first week, before the project kept a changelog. They are the build that got dictation working end to end: a hotkey, Whisper, and text in the window you were already in. They are on GitHub if you want them, but there are no release notes to quote, so this page starts at v0.1.6.