When Singer Maker splits a song, the whole job runs on the Mac in front of you. Nothing uploads and nothing queues. This post explains what that job actually is.
The problem, stated plainly
A mixed song is one waveform. The voice, the drums, the guitars — all of it has been summed into a single signal, and there is no algebra that neatly un-sums it. Instruments overlap in time and in frequency; a hi-hat and a consonant can occupy the same spectral neighborhood at the same instant.
So separation is not "undo the mix." It's a prediction problem: given this mixture, what is the model's best estimate of the vocal? Neural networks turn out to be very good at this, because they've seen the answer key. Separation models train on large sets of songs where the isolated stems are known, and they learn what voices do — how harmonics stack, how vibrato moves, how a breath differs from a cymbal wash.
What runs on your Mac
Singer Maker's primary engine is a BS-Roformer model — a transformer that works on the song's spectrogram, the time-frequency picture of the audio. The model reads the mixture in frequency bands, decides for every point in time and frequency how much of the energy belongs to the voice, and rebuilds two signals from that decision: a vocal stem and an instrumental stem. Exactly two, on purpose. For vocal practice, "the voice" and "everything else" is the split that matters, and doing one split well beats doing four splits half-well.
This class of model is heavy. Running it in reasonable time takes a GPU, and that's why Singer Maker is Apple Silicon-only: M-series chips share one pool of memory between CPU and GPU, so a several-hundred-megabyte model and a full-length song can sit in the same place while the GPU chews through them. We run the models through MLX, Apple's array framework built for exactly this hardware. An Intel Mac without that unified-memory GPU simply wouldn't get through a track in acceptable time, so rather than ship a bad experience, we don't ship to Intel at all.
The models themselves download once, on first run. That keeps the installer small, and it's one of the four honest network touchpoints the app has (the others: song imports you ask for, update checks, and license activation). Audio never goes the other way — nothing you import or record leaves the machine.
The part most tools don't tell you
Separation quality is not uniform. A dry lead vocal over a sparse arrangement separates beautifully. Heavy reverb tails, dense choir stacks, and distorted walls of guitar are harder, and no model, cloud or local, gets them perfect. You can sometimes hear faint vocal residue in an instrumental, or a touch of cymbal in a vocal stem.
Singer Maker's answer is to measure instead of promise. After every split, the app scores the result across several signal metrics and shows you an honest quality score per track. When a song is hard, you'll see it instead of wondering about it. And because your original file is never modified (splits live next to it, not in place of it), a re-split with a different engine mode costs nothing but time.
Why local wins for this job
For a practice tool, on-device separation isn't ideology; it's just the better engineering fit. The song is already on your disk. The GPU is already in your lap. Sending audio to a server adds a round trip, a payment meter, and a privacy question — and returns the same two stems you could have computed where the file already lives.
Your Mac has been able to do this for a while. Singer Maker just puts it to work. If you want to hear your own library split in two, the trial is the full app for 14 days.