Repository navigation
SoundOff v1.0.0
SoundOff turns a recording into an editable transcript on your own computer. No account, no upload, no hosted
inference. This is 1.0: the workflow it was built for — record or import, transcribe, listen and correct, export —
is complete on Windows. Read what each download can actually do.
New in 1.0
- Playback speed, with the pitch kept. 0.5× to 2× from the button beside the clock, or Ctrl+Shift+, and
Ctrl+Shift+. to step. Timing stays in the recording's own time at any speed. - One audio engine on every platform. Playback moved to SoundFlow over miniaudio, so the macOS and Linux builds
now play back and record too. - Recording on macOS and Linux. A microphone, or the computer's own sound where the system offers it: the sound
server's monitor on Linux, a loopback driver such as BlackHole on macOS. Recording both at once stays Windows-only. - Autosave. On by default: a pause in typing saves a revision of its own, without moving your cursor. Undo still
steps back one burst of typing at a time, and closing never asks you to throw work away. - Notes that say what happened. Saved, Copied, Exported, Recording added and the rest appear briefly at the
bottom of the window and go by themselves.
Since 0.2.0 also: the start screen's two ways in, the waveform drawn as a timeline of who spoke when, Alt+Up/Down
between paragraphs, and double-click to play from a point.
Which download
| Download | What it can do |
|---|---|
SoundOff-1.0.0-win-x64.zip |
Everything, and the only build that has been run: import, record (including microphone and computer together), transcribe, play back at any speed with the spoken word lit inside the text, correct, export. |
SoundOff-1.0.0-osx-arm64.zip, SoundOff-1.0.0-osx-x64.zip, SoundOff-1.0.0-linux-x64.tar.gz |
The same app, with playback and recording through the same engine as Windows. Never launched on a Mac or a Linux machine by anyone. Treat them as something to try. |
What every build needs
- ffmpeg and ffprobe on
PATH(or inSOUNDOFF_FFMPEG_DIR). They identify imported media, decode audio for
recognition and build playback proxies. Nothing that touches media works without them. - Transcription needs the private Python runtime, installed once with
python scripts/setup_runtime.py, and
then a one-time model-pack download of about 2 GB from inside the app. See the README. The runtime's paths are
correct on macOS and Linux, but transcription has never been run there. - Nothing else: the app is self-contained, demo included.
macOS
The macOS builds are unsigned and not notarized, and they are a plain executable rather than a .app
bundle. Gatekeeper will refuse to open them until you clear the quarantine attribute yourself, and they launch
from a terminal — which is also what macOS will ask about the first time you record from the microphone:
xattr -dr com.apple.quarantine SoundOff-1.0.0-osx-arm64
./SoundOff-1.0.0-osx-arm64/SoundOff.Desktop
Do that only if you are comfortable running an unsigned binary you have checked the provenance of.
Linux
tar xzf SoundOff-1.0.0-linux-x64.tar.gz
./SoundOff-1.0.0-linux-x64/SoundOff.Desktop
What was verified
On Windows, on real hardware: 379 tests passed, and scripts/verify.py --clean --desktop-smoke rebuilt from
clean with a locked restore and exercised every real adapter — WhisperX recognition and alignment, WASAPI loopback
capture, playback through SoundFlow on the real output device (2× measured at 2.00×), and the new macOS/Linux
recorder run over miniaudio's loopback on this machine's audio stack. docs/VERIFICATION.md records the commands,
the numbers and the limits, including what is not covered: no Mac or Linux run at all, no listening test of
the time-stretch, no screen-reader or keyboard-only pass, and no performance claim.
The release workflow runs the same test suite on a hosted Windows runner minus the suites that need a real audio
device or the model pack, which a hosted runner does not have.
MIT licensed. Playback and portable recording use SoundFlow (MIT) and miniaudio (MIT or Unlicense). The
repository's code is AI-generated; the README says so at the top.