Moosic

Overview

Moosic is a vocal training app for macOS. Drop in a song and it pulls the voice out, lines the lyrics up in time, and grades your pitch against the original while you sing.

Three models do the work and each is doing a job the others cannot. Meta's Demucs handles source separation, so you get a clean vocal reference and an instrumental to actually sing over. Whisper transcribes the lyrics with word-level timestamps, which is what makes real-time scoring possible — without per-word timing you can only grade a whole take after the fact. Pitch tracking runs through Apple's Accelerate framework rather than a Python DSP stack, because the feedback has to be immediate and a round trip through anything heavier is audible as lag.

All of it runs on the machine. There is no upload step and no account, partly on principle and partly because singing badly in private is the entire use case. Apple Silicon is strongly recommended since Demucs and Whisper are doing real work on every import.