Can a prompt replace Moises?

Build a local practice tool for musicians around source separation.

Separation runs through Demucs — do not attempt to train or implement a separator. Take an audio file, produce vocals, drums, bass and other as separate WAVs into a folder named after the track, and cache them so a song is separated once.

The practice interface is the actual product: a mixer with one fader and a mute per stem, so removing the guitar to play along is one click. Add independent tempo and pitch control using a phase vocoder — slowing a passage to 70% without shifting pitch, or transposing to fit your instrument, are the two things a practising musician wants most.

Add a loop bar: drag to select a region, loop it, and let a keyboard shortcut nudge the boundaries by a beat. Practising a four-bar fill means playing it forty times, so getting the loop points comfortable matters more than any audio feature here.

Detect key and tempo with Essentia and print them, but display them as estimates rather than facts — automatic key detection is wrong often enough on real recordings that presenting it as certain will confuse someone practising.

Save a per-song state file with your fader positions, loop points and tempo setting, so reopening a song resumes the practice session rather than resetting it.

Do not build chord transcription. Chord recognition on dense mixes is unreliable, and a wrong chord chart is worse than no chart when you are learning a song by ear anyway.