SoundOwl's icon: an owl under a crescent moon

Make voices clear again.

Phone lines, cheap microphones and heavy compression throw away the top half of a voice. SoundOwl rebuilds it on your Mac, live as you listen and in your recordings.

Download for Mac
brew install --cask fanckush/soundowl/soundowl

Free and open source. Works offline. macOS 15 or later on Apple Silicon, version v0.1.0.

Restored

0:00
A public-domain LibriVox reading of Alice in Wonderland, degraded on purpose, then restored by SoundOwl. On a keyboard, hold Space.

An owl in your menu bar.

Click it and everything SoundOwl does is right there. It's made for speech, and it leaves everything else alone.

Calls and videos

Switch on the app you call or watch with: Zoom, Teams, a browser. Voices are restored as you hear them, about a fifth of a second behind. Music and sound that's already clear play untouched. Headphones only, unless you say otherwise.

Clear copies

Drop recordings on the owl, or right-click them in Finder and choose Restore with SoundOwl. Copies appear next to the originals with their title, artist, album and artwork. A Shortcut and soundowl restore in Terminal do the same.

Listening

Open a podcast, lecture or audiobook with SoundOwl and it plays restored right away, even if it's ten hours long. Hold Space to hear the original, let go to hear the difference. It remembers where you stopped.

The SoundOwl panel open under its owl in the menu bar, with QuickTime Player switched on and its options showing, next to the player window restoring a sample recording. The waveform is blue up to the playhead.

How it works

SoundOwl runs LavaSR v2, a 50 MB speech restoration model by Yatharth Sharma (paper). It looks at the part of the voice that survived and predicts the frequencies that didn't, up to 48 kHz. The optional noise reduction is UL-UNAS by Xiaobin Rong and colleagues.

The model downloads once, the first time you open the app, and then everything happens on your Mac. Your recordings are never uploaded. The models run in ONNX Runtime; the signal processing around them is native Swift on Apple's Accelerate framework, tested to match the reference implementation.

Long recordings are restored in 30-second pieces, starting from where you're listening, so you never wait for the whole file before you can hear it. Live works on the sound as it arrives, in small overlapping blocks, and Apple's built-in sound classifier tells it when an app is playing music, so it steps aside. The engine notes have the details.

Speed
About 140× realtime on an M5 Pro. An hour of speech in under 30 seconds.
With noise reduction
About 48× realtime.
Live
About 0.2 s behind the app. Faster or clearer per app.
Model download
52 MB, once. Credited and removable in Settings.
Formats
Opens MP3, M4A, AAC, WAV, AIFF and FLAC. Exports AAC, Apple Lossless or WAV.
Your audio
Stays on your Mac.

Free, and it stays that way

SoundOwl is open source under the Apache 2.0 license, the same license as the model it runs. Read the code, file an issue, or send a pull request on GitHub. If it saves you an afternoon, you can sponsor its development.