Add an ML‑based “Vocal ⇆ Instrumental” Source‑Separation Layer

[Sam-Mucyo]

✨ Motivation

After a few tests with stereo audio, we found most of them don’t come with discrete vocal and instrumental stems.
By adding an ML source‑separation layer we can:

• Dynamic mixing – route isolated vocals to the center channel, send drums/bass to LFE, etc.

• Karaoke / practice mode – solo or mute vocals on‑the‑fly.

• Accessibility – raise vocal intelligibility for listeners with hearing loss.

🛠 Proposed Approach

Item	Detail
Model	Start with the open‑source Hybrid Transformer Demucs v4 checkpoint (MIT‑licensed, PyTorch).
Inference mode	Offline (full track) first; near‑realtime streaming as a follow‑up.
Interface	New CLI flag --separate=2 or --separate=4, plus player UI toggle.
Output	Standard 32‑bit float WAV stems: vocals.wav, accompaniment.wav (or drums/bass/other).
Routing	Default mapping: vocals → center + height, accompaniment → L/R + surrounds. Configurable in config.yaml.

📚 Below References Might be helpful

• Demucs v4 paper & code – https://arxiv.org/abs/2211.15662

• MUSDB18‑HQ dataset – https://sigsep.github.io/datasets/musdb.html

• nussl toolkit – https://github.com/interactiveaudiolab/nussl