Feature index
This is the complete inventory, not a highlight reel. Every entry below states plainly whether it is in the release you can download today or whether it only exists in the development branch.
01
Hold a global hotkey — Left Ctrl + Left Alt by default — speak, and release; the text lands in whatever window had focus. It is registered with the OS, so it fires in editors, terminals and reply boxes alike. Fully rebindable.
Double-press the same combo within 400 ms to lock recording on; press again to stop.
An always-on-top 42-pixel capsule that expands only while listening. Shows state, a live waveform, and a right-click menu with the mode selector and quick toggles.
Words appear in the pill while you are still speaking. Opt-in.
Hides the pill for screen recordings and screenshots while keeping it interactive. Recoverable from the tray.
An RNNoise pre-pass strips fan noise, keyboard clatter and non-speech before transcription.
Lowers system volume while you dictate so music does not bleed into the mic, then restores it. Default 70%.
Mic audio is downsampled to 16 kHz through a windowed-sinc filter rather than linear interpolation.
Choose which microphone OmniVox listens to.
Starts with Windows via the registry Run key.
Light is a warm-paper treatment.
02
Clipboard (copy only), Type (a clipboard-verified paste that then restores whatever you had on your clipboard), or Both.
Presses Enter after the text lands, so a dictated message sends itself. Type and Both only.
Formal, Casual or Very Casual capitalisation and punctuation, overridable per context mode.
Drops “um”, “uh”, “you know” and stutter repeats. Turn it off for verbatim quoting.
03
Spoken phrases that become real keystrokes mid-dictation. Every command has its own on/off toggle, a rename field, and a trigger scope of Anywhere or At end.
“new line”, “new paragraph”, “delete last word”, “select all”, “copy that”, “cut that”, “undo that”, “redo that”, “press tab”, “press escape”, “press enter”.
“bullet point” and “next bullet” produce dash-prefixed items; “number item” auto-counts 1. 2. 3.; “end list” closes it. The list lands as one atomic paste.
Saying “First… Second…” renders a numbered list with the ordinal words stripped.
As the final word, presses Enter. Independent of Ship Mode.
Mouse click, right click, double click, scroll up, scroll down and switch window. Off by default.
Map your own phrase to a key combo, a mouse action or an app launch.
04
A separate dedicated hotkey — Right Ctrl by default — where the whole utterance is an instruction rather than dictation. Off by default.
“open Spotify”, “switch to Slack”, backed by a scanned app index.
“search the web for…”, “go to…”. A spoken URL must be grounded in what you actually said; anything ungrounded parks on the confirm pill.
Copy, paste, save, new tab, close tab, screenshot, refresh, find, next and previous tab, and more.
Play and pause, next and previous track, mute and unmute, volume up and down. Unmute is deterministic rather than a blind toggle.
Minimise, maximise, close, full screen, minimise everything.
“type …” pastes for review; “tell Claude to …” pastes and presses Enter. Every send routes through the confirm pill.
One utterance can chain steps: “open Spotify and play it”.
Enter confirms, Esc dismisses, Ctrl+Enter sends an editable message, from an overlay that never takes focus.
Aborts a running command, or reverses the last action: re-closing a launched app, restoring a minimised window, undoing a non-submitted paste.
Every side-effecting action is bound to a specific window handle and owning process, and re-verified as foreground before it fires.
05
Bundles of writing style, dictionary, snippets, vocabulary, app bindings and a Structured Mode profile, saved together. The built-ins are General, Programming, Business & Sales, Medical and Legal, each seeded with a domain vocabulary. Custom modes are unlimited.
The active mode follows the foreground application when recording starts.
06
Add a word like “OmniCue” and a deterministic post-transcription pass rewrites anything that sounds like it — same consonant skeleton, edit distance within 3 — to your exact spelling. Split compounds fuse back together.
“Heard as → Replace with” rules, seeded with common corrections.
A trigger word expands to longer text. Global, or scoped to one mode.
Opt-in. Reads visible text from the focused window through Windows UI Automation, extracts identifiers, file paths and CLI flags, and feeds them to Whisper so they transcribe verbatim. Never leaves the device.
07
A fully local LLM pass that reshapes a raw dictation into a finished document, held to shape by a GBNF grammar so the output is physically schema-valid.
Agent Prompt (goal, context, constraints, files, urgency), Email (recipient, subject, body, sign-off) and Notes (a hierarchical outline). One per context mode.
Optional: structure only the utterances you end with “voxify”. Fifteen-plus phonetic variants are accepted, so a Whisper misread still triggers.
Output is grounded in the raw transcript: files must appear in what you said, slots are de-duplicated, and third-person is rewritten to first.
A 420-pixel panel with the Markdown preview, metadata chips, a raw-transcript drawer, and Edit, Paste, Copy, Raw and Dismiss actions. “Raw” pastes your exact pre-LLM words.
On timeout, a missing model or a parse failure it falls back to plain text rather than failing the dictation.
The prompt cache is held across dictations, released after five idle minutes and rebuilt in the background the moment you press record.
08
From Tiny (75 MB) through Medium, Medium Q5, Large V3 Turbo and Distil Large V3 (1.52 GB, roughly five times faster than large-v3). English and multilingual builds.
The multilingual builds transcribe with automatic language detection, or translate into English on the way out.
OmniVox measures the machine and recommends a model tier rather than asking you to guess, and shows the active backend.
Qwen3 1.7B Instruct Q8_0 (1.83 GB, the default for Structured Mode) and Qwen3 0.6B Instruct Q8_0 (639 MB).
One build flag covers both Whisper and the LLM, on AMD and NVIDIA alike. Roughly three to five times faster than CPU.
If the GPU fails to load, the overlay says it is running on CPU instead of silently degrading, and every model load is logged with its backend and duration.
09
Recording a call end to end — both sides of the audio, a live transcript, and notes that cite it — on the same local Whisper engine dictation already uses.
Not in v0.6.3
Built and running in the development branch. None of it is part of v0.6.3, and none of it is in the download this page links to.
Two independent audio streams: your microphone, and Windows system output via WASAPI loopback. Works with Meet, Zoom, Teams, Webex, Discord or anything else playing audio.
The same Whisper engine and model as dictation. Audio is sealed to disk in recoverable 20-second chunks, and each chunk is deleted the moment it has been transcribed.
Process everything on stop so Whisper never competes with your call, or transcribe live as chunks seal.
Microphone segments are labelled “You” and system-audio segments “Meeting”. Rename any speaker and apply it to every segment from that source at once. This is source-based attribution, not neural diarization.
Polls for two independent signals — the Windows microphone consent store plus a window-title scan — before suggesting a recording. Gated behind a setting.
A draggable dock with a voice-reactive orb, dual channel meters, an elapsed timer, and pause, resume and stop.
Type or dictate rough notes during the call, beside a searchable live transcript with timestamp jumping and per-segment editing.
Alt+M drops a timestamped marker the summariser treats as an importance signal.
Overview, topics, decisions, action items, open questions and highlights, every item carrying an [HH:MM:SS] reference back to the transcript. Invented references are stripped by a local validator before anything renders.
General notes, Executive brief, One-on-one, Sales call, Interview and Standup, plus free-text custom instructions.
Grounded follow-up questions. Retrieval runs locally: only the most relevant segments are selected on your machine, and the panel tells you how many of how many were used.
Durable action items that survive note regeneration, with a cross-meeting inbox, owner filters, due dates and jump-to-source.
The last fifty note snapshots, with hand-edits captured automatically and safe restore.
Carry agenda, participants and open action items forward from the last meeting with the same title.
Reusable meeting setups, full-text search across titles, notes and transcripts, and a restorable trash.
Sealed audio is resumed after a crash, and a potentially billed AI request is never silently repeated.
Polished Markdown notes, a full Markdown archive, a full JSON archive, WebVTT and SRT.
10
An auto-saving editor in the main window; dictation inserts straight into the open note.
A small always-on-top pad you can dictate into from anywhere, in Cards or Note view, with a voice-reactive orb and a capture toggle that routes your hotkey into the pad while you read elsewhere. Remembers size and position, and survives a restart.
A searchable local archive of every dictation, with delete and export.
Lifetime words, characters, dictations and time recorded; day streaks, average pace and active days; a contribution heatmap, a peak-hours histogram and a 30-day trend. Computed locally — the full transcript text never crosses the IPC boundary.
11
Dictation audio and transcription are entirely on-device under every configuration. Audio is never uploaded.
There is no server to sign into, and no analytics SDK, crash reporter or auth library in the dependency lists.
History, notes, dictionary entries and modes live in omnivox.db in your own app-data folder. Delete a dictation and it is gone from disk.
Fail-closed on the save path, and existing rows are securely purged.
New installs keep 30 days by default, adjustable up to forever, with a bounded cleanup job every 24 hours.
Every internal command declares which of the app’s windows may call it.
Meeting AI notes are the one feature that can leave the machine, and only if you supply your own OpenRouter key. It is inactive until you do. Audio is never sent — only transcript text and the notes you typed. The key is stored in Windows Credential Manager, and a monthly spend guardrail (default $2) and a per-meeting cap (default $0.05) are enforced locally before any request.