Creative Tooling
Opus Vela Voice
Speak anywhere on Windows and keep every word on your own machine — the recognition, the transcripts and the history all stay on the device.
Outcome
A local Whisper runtime turns speech into text and places it straight into whichever application has focus — checking each target is safe to write to, verifying the result wherever accessibility APIs allow it, and keeping a local recovery history for the times insertion fails.
Media
Select any frame to enlarge
Live Voice product page
The current Voice product page: local recognition, recoverable history and write-safety for dictation across Windows applications.
Current application build
The current underlying application build, captured before the interface-label migration to Opus Vela. Every dictation is written to local history before insertion is attempted, so it remains recoverable when a target application cannot accept it. The mark is the waveform the Voice Bar draws while it is listening. It is the icon Windows shows in the taskbar, in Alt-Tab and on the executable itself. The Voice Bar while the hotkey is held. The waveform is the real microphone level, so it is visible whether the machine is actually hearing anything. Usage analytics are content-free by construction: counts and categories only, never transcript text. An automated test fails the build if a dictated canary string reaches the analytics table. Thirty languages, each shown in its own script. The list is what the local model handles well rather than everything it claims, because a language it transcribes badly is worse than one not offered at all. The settings that decide how a dictation behaves, in one pane: the hold shortcut, which microphone, which language, and whether every transcript is also copied to the clipboard. Voice styles match tone to the application being dictated into. Built-in profiles preview their exact changes; custom instructions fail closed until a provider exists rather than silently doing nothing. Transforms rewrite a dictation after the fact. Each previews its plan, its protected placeholders and the revision it expects, so a rewrite is inspected before it is trusted. Quick Fix keeps a recent dictation recoverable and correctable without sending the transcript to a hosted transcription service. Light and dark follow the Windows theme.
Case study
The problem
Dictation tools either ship audio to someone else's cloud or trust every text field equally. Opus Vela Voice is built on the opposite assumptions: audio is process-lifetime-only, and no target application is written to until it has been checked as safe to write to.
Architecture
Two local processes. An Electron hub holds the SQLite-backed application service, the transcript pipeline and the interface; a native C# companion owns what the Electron sandbox cannot reach — the system-wide keyboard hook, raw WASAPI audio capture, and UI Automation text insertion into other applications. They speak only over authenticated loopback. Recognition runs against a quantised Whisper model kept resident in memory, because loading it per dictation is the difference between usable and not.
The hard part nobody plans for
Latency was erratic rather than merely slow, and the cause was not the model. On a four-core laptop the same request measured 3.7 seconds idle and 28 seconds while other applications saturated the CPU. Quantisation and sizing the audio context to the actual speech helped, but the decisive change was scheduling: recognition is interactive work the operator is waiting on, and had to be prioritised as such. Getting there meant building an automated benchmark first, because single manual timings were hiding an eight-fold variance.
Where it is now
Live. The Windows build is downloadable from opusvela.com and the application is in daily use. Voice was relicensed from MIT to the Elastic License 2.0 in August 2026 and its repository is now private, so it is source-available on request rather than open source. Dictation reaches focused Windows controls rather than a fixed application allow-list, and reports honestly when it cannot read text back. GPU-accelerated recognition remains open distribution work.
Verified architecture
System map
Technology
- TypeScript
- C#
- Electron
- SQLite
- whisper.cpp
- Windows UI Automation
- WASAPI
Scope and boundaries
Voice is source-available under the Elastic License 2.0 with a private repository; source access is arranged on request. Recognition quality is bounded by the local model, GPU acceleration is not claimed and translation is into English rather than between arbitrary language pairs. Audio never needs a hosted transcription service.