Overlay stays stable across sessions
The floating bubble survives window manager changes, display sleep, and app switching.
Qt6 window flags ensure always-on-top behavior. Position persists across recording sessions.
Measured Reality.
Documented performance, hard limits, and architectural constraints. Every claim is dated, route-linked, and supported by actual measurement.
Measured under sustained load, not averaged across runs.
Behavior verified across versions, not just point releases.
Transcript history preserved
Every transcription is logged locally with timestamp and full text.
History window shows all past transcriptions. Copy, review, or clear at any time.
The full range of real-world conditions, not just the demo.
Verified per platform. Not implied, not extrapolated.
Cross-platform, on-device everywhere it ships
One Lven Instant subscription covers every platform as it ships — Windows, Linux, and Android today, macOS and iPhone on the way. Every shipped build runs the same on-device transcription engine.
Each platform uses its native audio stack and text injection method. No emulation layers, no shared web shells.
Measured in state accuracy, output integrity, and traceability.
Text lands at cursor position
Transcribed text is injected into the active input field at the exact cursor location.
On Windows, Lven tries UI Automation and falls back to a controlled clipboard paste. Linux currently uses xdotool on X11 or wtype on Wayland. macOS and iPhone remain unshipped, and Android uses its current platform service.
Documented before failure, not after.
Local-first recording
Audio is written to disk before transcription. Network failure does not lose the recording.
WAV files persist in the cache directory until transcription succeeds or the user clears them.
Engineered into the system, not retrofitted.
Transcription is instant
Speech becomes text on your device with no upload step and no network round trip. The result appears the moment you stop speaking.
The quantised speech model is downloaded during setup and then runs on your CPU. No remote inference means the transcription path has no upload queue, server rate limit, or network round trip.
Documented architectural limits, not aspirational targets.