Skip to content

Releases

SoundOff Android 0.5.0

Pre-release

Choose a tag to compare

@github-actions github-actions released this 16 Oct 12:48

SoundOff for Android: on-device recording and transcription with English, Filipino and mixed-language modes.

  • Stop and Resume append to the same interview; transcription catches up independently.
  • Waveform scrubbing, playback speed controls, and transcript editing.
  • Existing-audio import, WAV/MP3 export, and clipboard copy.
  • Multilingual Tiny, Base and Small Q5 models download separately in the app.

Install the attached APK on Android 8.0+ ARM64 devices (x86-64 is also included). Check the attached verification report for the exact tests and limitations of this version. This is the Android release stream; desktop installers are published under v* tags.

Updates require the same signing key. A release-signed APK cannot replace a debug-signed installation with a different key. Export recordings before uninstalling any existing installation; uninstalling deletes its private recordings.

SoundOff v1.1.0

Choose a tag to compare

@github-actions github-actions released this 06 Oct 04:45

SoundOff turns a recording into an editable transcript on your own computer. No account, no upload, no hosted
inference. 1.1 makes it installable: a real installer on every platform, and transcription set up from inside the
app instead of from a terminal. Read what each download can actually do.

New in 1.1

  • Installers. A Windows installer that needs no administrator rights, a Mac disk image, and a Debian/Ubuntu
    package. They replace the zip files.
  • ffmpeg comes with the Windows installer. The Debian package installs it for you. On a Mac it comes from Homebrew,
    which the app now finds even when it is opened from the Finder.
  • Set up transcription from the app. A button in the Transcribe card installs the speech engine in its own
    window, and SoundOff notices when it is done without a restart. The Windows installer offers it on its last page.
    No Python install or repository download needed.
  • Transcription setup on macOS and Linux now installs where the app looks. It used to install somewhere the app
    never checked.

Which download

Download What it can do
SoundOff-…-win-x64-setup.exe 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. Installs for you alone, with no administrator prompt, and carries its own ffmpeg.
SoundOff-…-osx-arm64.dmg (Apple Silicon), SoundOff-…-osx-x64.dmg (Intel) The same app, with playback and recording through the same engine as Windows. Never launched on a Mac by anyone. Treat it as something to try.
SoundOff-…-linux-x64.deb (Debian, Ubuntu), SoundOff-…-linux-x64.tar.gz (anything else) The same again. Never launched on Linux by anyone.

Setting up transcription

Transcription needs a speech engine that is installed once, after the app: about 2 GB on disk. The Windows
installer offers to do it on its last page. Otherwise press Set up transcription in the app's Transcribe card,
which opens a window that shows the download and says when it is done. Then Prepare model pack in the app
downloads the speech models, about 2 GB more, also once. The engine has never been set up on a Mac or on Linux.

macOS

Open the disk image and drag SoundOff onto Applications. The app is not signed or notarized, so macOS will
refuse to open it the first time. Open it once, then go to System Settings → Privacy & Security and choose
Open Anyway. Do that only if you are comfortable running an unsigned app you have checked the provenance of.

The Mac app does not carry ffmpeg. Install it with Homebrew: brew install ffmpeg.

Linux

sudo apt install ./SoundOff-…-linux-x64.deb

That installs ffmpeg too, and puts SoundOff in the applications menu. With the archive instead, install ffmpeg
yourself, then run SoundOff.Desktop from the unpacked folder.

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.

SoundOff v1.0.0

Choose a tag to compare

@github-actions github-actions released this 05 Oct 07:30

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 in SOUNDOFF_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.

SoundOff v0.2.0

SoundOff v0.2.0 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 04 Oct 16:19

SoundOff turns a recording into an editable transcript on your own computer. No account, no upload, no hosted
inference. This is a prerelease: read what each download can actually do.

New in 0.2.0

  • Two ways in. The start screen offers Transcribe a file and Record side by side, with the recorder
    right in the Record tile. Recent projects open with one click and say when you last opened them.
  • The next step where you are looking. A project with audio and no transcript offers Transcribe in the
    middle of the window, with its progress and Cancel in the same place.
  • The transcript on the waveform. Each paragraph is a band in its speaker's colour, the one being spoken is
    lit, and there is a time ruler and a readout of the time under the pointer. Double-click to play from a point;
    Shift+wheel or a sideways swipe slides along the recording.
  • Recording you cannot lose track of. A red indicator in the top bar shows a running take's clock and Stop,
    even with the side panel hidden.
  • Moving around. Alt+Up and Alt+Down step between paragraphs; double-clicking a paragraph's timestamp plays
    from it. Dragging a file over the window shows where it will go.
  • Quieter chrome. The top bar only shows what applies to what is open, and a dot beside the project's name
    marks unsaved changes.

Which download

Download What it can do
SoundOff-0.2.0-win-x64.zip Everything: import, record, transcribe, play back with the spoken word lit inside the text, correct, export.
SoundOff-0.2.0-osx-arm64.zip, SoundOff-0.2.0-osx-x64.zip, SoundOff-0.2.0-linux-x64.tar.gz Import, edit, export, bundles. No playback and no recording — those adapters are Windows-only and the app says so where their controls are.

Only the Windows build has been run. The others are built from the same source by the same workflow and
have not been launched, on any machine, by anyone. Treat them as something to try, not something to rely on.

What every build needs

  • ffmpeg and ffprobe on PATH (or in SOUNDOFF_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, and New demo project… starts its fixture worker through the
    worker's own launcher, so it does not need a .NET install either.

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:

xattr -dr com.apple.quarantine SoundOff-0.2.0-osx-arm64
./SoundOff-0.2.0-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-0.2.0-linux-x64.tar.gz
./SoundOff-0.2.0-linux-x64/SoundOff.Desktop

What was verified

On Windows, on real hardware: 358 tests passed, and scripts/verify.py --clean --desktop-smoke rebuilt from
clean with a locked restore and exercised the real inference, capture and playback adapters — real WhisperX
recognition and alignment, real WASAPI loopback capture, and the real output device. The window was run and
looked at in both themes and both views. docs/VERIFICATION.md records the commands, the numbers and the limits,
including what is not covered: the new drop target and recording indicator were not seen during a real drag
or a real take, there is no screen-reader or keyboard-only pass, no performance claim, and no macOS or Linux run
at all.

The release workflow runs the same test suite on a hosted Windows runner minus the three suites that need a real
audio device or the model pack, which a hosted runner does not have.

MIT licensed. The repository's code is AI-generated; the README says so at the top.

SoundOff v0.1.0

SoundOff v0.1.0 Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 21 Sep 07:21

SoundOff turns a recording into an editable transcript on your own computer. No account, no upload, no hosted
inference. This is the first tagged build, and it is a prerelease: read what each download can actually do.

Which download

Download What it can do
SoundOff-0.1.0-win-x64.zip Everything: import, record, transcribe, play back with the spoken word lit inside the text, correct, export.
SoundOff-0.1.0-osx-arm64.zip, SoundOff-0.1.0-osx-x64.zip, SoundOff-0.1.0-linux-x64.tar.gz Import, edit, export, bundles. No playback and no recording — those adapters are Windows-only and the app says so where their controls are.

Only the Windows build has been run. The others are built from the same source by the same workflow and
have not been launched, on any machine, by anyone. Treat them as something to try, not something to rely on.

What every build needs

  • ffmpeg and ffprobe on PATH (or in SOUNDOFF_FFMPEG_DIR). They identify imported media, decode audio for
    recognition and build playback proxies. Nothing that touches media works without them.
  • The .NET 8 runtime on PATH for New demo project… only. The app itself is self-contained; the demo's
    fixture worker is a separate process the app starts with dotnet.
  • 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 as of this build, but transcription has never been run there.

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:

xattr -dr com.apple.quarantine SoundOff-0.1.0-osx-arm64
./SoundOff-0.1.0-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-0.1.0-linux-x64.tar.gz
./SoundOff-0.1.0-linux-x64/SoundOff.Desktop

What was verified

On Windows, on real hardware: 354 tests passed, and scripts/verify.py --clean --desktop-smoke rebuilt from
clean with a locked restore and exercised the real inference, capture and playback adapters — real WhisperX
recognition and alignment, real WASAPI loopback capture, and the real output device. The window itself was run
and measured rather than only tested. docs/VERIFICATION.md records the commands, the numbers and the limits,
including what is not covered: no screen-reader or keyboard-only pass, no performance claim, and no
macOS or Linux run at all.

The release workflow runs the same test suite on a hosted Windows runner minus the three suites that need a real
audio device or the model pack, which a hosted runner does not have.

MIT licensed. The repository's code is AI-generated; the README says so at the top.