Daniel Murrell released OnKey 0.1.0 on October 22, a free, open-source Mac singing trainer that learns a song's melody from music playing in another app. It lets you practise against the original singer's notes using Spotify, Apple Music or YouTube playback, without OnKey supplying the music itself.
The first public release processes songs on your Mac and draws your pitch against the learned notes on subsequent plays. There are practical limits: the chart needs an initial listening pass, song identification can involve Shazam, and a copy you build yourself cannot use that Shazam connection. This report covers the release and its documentation. We haven't run the Mac app or verified its pitch display or scoring.
The second play gets the note chart
OnKey needs Apple Silicon, starting with M1, and macOS 14.2 or later. Setup requests microphone access to hear you and system audio access to hear the music app. You can allow it to follow Spotify and Apple Music, or choose Listen only to follow songs by sound, according to the versioned README.
The first play shows lyrics while OnKey gathers the song. After playback ends, a model separates the lead vocal, CREPE tracks the singer's pitch and the GAME model converts the melody into note bars. Once learning is complete, later plays show those bars with your voice drawn over them and a report for each line. That sequence makes the documented workflow better suited to repeating a song than opening an unfamiliar track for an immediate note-by-note exercise.
The score is deliberately forgiving. The scoring rules give full pitch credit within half a semitone and accept singing an octave above or below the original. Each half-second can shift by up to 80 milliseconds to accommodate timing differences. The overall score also accounts for how much of the singer's part you sang, so a brief accurate phrase doesn't earn the same result as sustained participation.
Those rules have limits. Murrell's scoring documentation says rap and spoken passages barely suit pitch scoring, while harmonies can make the app follow the wrong voice. Loud speaker playback can also remain in the microphone signal after echo cancellation and be mistaken for singing. Headphones are the project's recommended route to more reliable scores.
Local analysis still makes network requests
The privacy documentation says OnKey analyses music in memory and never saves or uploads the music audio or separated vocal. It keeps learned notes, timing information and microphone recordings on the Mac. A microphone recording made while using speakers can still contain some of the song.
Song lookup has a separate network footprint. When no music app names an unfamiliar song, OnKey can send Shazam a signature derived from up to five seconds of the music app's sound. The project says this is neither the audio itself nor microphone input. It also sends song metadata to LRCLIB for lyrics and Apple services for covers and track information, and contacts GitHub for updates. The privacy page documents switches for Shazam, lyrics and automatic update checks.
A self-built copy loses Shazam
OnKey's code is MIT-licensed, and the build instructions list an Apple Silicon Mac, Xcode command-line tools, Rust, Node and Python 3. But Murrell says his Apple signature is needed to ask Shazam to identify a new song. A self-built copy therefore loses that feature. Spotify and Apple Music can still report their own track names, so the limitation affects YouTube and Listen only use.
The next useful check is how consistently the app follows less straightforward recordings. The documented limits already identify harmonies and speaker leakage as failure cases. Testing those against the signed release, alongside the reduced song-identification options in a self-built copy, would show how far the published workflow holds up beyond clear lead vocals and headphones.