OnKey is a solo-developer passion project that does something deceptively simple: it listens to whatever song you're playing in Spotify, Apple Music, or YouTube, learns the melody on first listen, and then scores your singing on every subsequent play. The pitch tracking scrolls toward you like a rhythm game, your voice is drawn over the target notes, and you get a line-by-line report. It's karaoke with a conscience — no music is stored, no audio leaves the machine, and the entire thing runs locally on Apple Silicon. The technical stack is genuinely interesting. On first play, OnKey separates the lead vocal using UVR's MDX-Net Karaoke 2, tracks pitch with CREPE (a neural pitch estimator), and converts the result into a per-syllable note map using the GAME note model. On subsequent plays, it matches your live session to the learned melody by listening to the music itself — no timecode sync, no metadata dependency. It forgives singing an octave down and doesn't penalize slight latency. These are small design choices that show someone who actually sings and knows what's annoying about pitch trainers. The privacy architecture is the standout feature and the thing most likely to earn trust from a skeptical user base. There's no account. No analytics. No crash reporting. The only outbound requests are Shazam signatures (not audio) for song identification, LRCLIB for lyrics, iTunes Search for cover art and metadata, and a daily GitHub update check — all of which can be toggled off. Audio never leaves the device. Takes stay on your Mac. The README even lists every outbound web address in a release. The fair-use posture is carefully drawn. OnKey doesn't provide, stream, or download music — it just listens to what's already playing. The developer explicitly asks users not to share the learned melodies, to use the built-in Share feature (which strips the music and exports only your pitch line and voice), and to use licensed karaoke services for public events. The lyrics come from LRCLIB, a free community database. This is a developer who has thought about the legal surface area and is trying to stay well inside it. The build-from-source option is a nice touch for trust. Clone the repo, install Rust, Node, and Python, fetch the models (SHA-256 verified), and build with Tauri. The only feature you lose is Shazam song identification, which requires the developer's Apple signature. Releases are signed with an Apple Developer ID and notarized. For a free tool with no revenue model, this is an unusual level of care. Limitations are real: Mac-only, Apple Silicon only, macOS 14.2+. That's a narrow audience. There's no Windows or Linux port, no mobile version, and no indication one is coming. The app depends on your music subscription for the actual audio — it's a layer on top of your existing setup, not a replacement for anything. And the scoring, while forgiving, is only as good as the CREPE pitch model and the initial melody extraction, which means some songs will inevitably be learned wrong. This is a tool for people who sing at home and want to get better at it without paying for a subscription service or uploading their voice to a server. It's not trying to be a social platform, a music service, or a business. It's a well-made instrument for a specific kind of practice, built by someone who clearly uses it himself.
