OnKey learns a song's melody from one play in Spotify or YouTube, then scores your singing on a Mac

Karaoke usually means somebody else’s catalog of songs. OnKey skips the catalog and learns the tune from whatever your music app is already playing. Its tagline: “Saving shower singers from themselves since 2026.”

The first time through a song, you get the lyrics. When the song ends, OnKey goes to work on your Mac. Per the README, “a model separates the lead vocal from the music, CREPE tracks the singer’s pitch, and the GAME note model turns it into one bar per syllable.”

Play it again and the notes scroll toward you. Your voice gets drawn over them, live, and you get “a score with a report on every line.” OnKey finds its place in the song from the music itself.

It doesn’t touch the music files. “OnKey doesn’t provide, stream or download music. Your own music app plays it,” the README says, and it “never saves the music it listens to.” The catch is hardware: you need an Apple Silicon Mac and macOS 14.2 or later.

What nobody has measured yet is how it holds up on the songs it admits are hard. The scoring page says pitch scoring “barely applies” to rapped or spoken songs, and with harmonies “some bars may follow a harmony.”

Why it matters

The grading is built for beginners. Every 10 ms counts. Land within 50 cents, half a semitone, and you get full credit, falling to nothing at 150 cents. “An octave above or below the singer counts as on the note, so you can sing in your own range,” the scoring page says, and each half-second can shift up to 80 ms, “so being a touch early or late doesn’t cost you anything.” Murrell says the rules are “set for beginners (I’m one).”

You can’t game it by humming one chorus. Your score scales with how much of the singer’s part you sang. “So one good line won’t get a good score, and someone who just listens scores close to nothing.”

He checked his own homework. Murrell ran 20 expert-rated recordings through it. OnKey’s “notes hit” number agreed with the experts at 0.77, against 0.88 for the experts agreeing with each other. He calls it “a good sign more than proof.”

The fine print on fair use is unusually plain for a hobby project. The README asks you to keep the melodies it learns to yourself, share takes through its own Share button (a video of your voice and pitch line, no music), and use a licensed karaoke service for a bar or an event. Its lyrics come from LRCLIB, a free community database.

Steal this

Before you install anything that asks for “system audio recording,” read what it sends out. OnKey lists every outbound request in its README: a Shazam signature (not the audio, never your mic), song details to LRCLIB, iTunes lookups and a GitHub update check, each with a switch in Settings. Kind of like checking which doors a house-sitter has keys to.
Practice with headphones on. The README warns that a take sung over speakers can contain some of the music, which matters the moment you hit Share. Same reason you don’t take a phone call next to the stereo.
If you’d rather not trust a stranger’s binary, build it yourself. The repo’s build steps fetch the models and check them by SHA-256, and spctl -a -vv /Applications/OnKey.app shows who signed a release. This is like reading the label instead of trusting the box.

The bigger idea is the one-play learn. If a laptop can turn any song you can legally stream into a scored singing lesson, the karaoke catalog stops being the product. What to watch is whether “stricter modes for better singers,” which the scoring page says are next, make it a real vocal coach, and how the labels feel about a melody transcriber that sits beside Spotify. Murrell is on X as @itsdsmurrell.

More Software