Stem Reconstruction - "Espresso" Client Project

The client had a track - "Espresso" - that needed to come apart before it could come back together: clean, production-ready stems separated from a mixed source, restored, and organized for further production work.

A full stem-separation and restoration pipeline, not a single-pass AI split:

First pass: Demucs (htdemucs_ft model) via UVR5 for initial full-band stem separation

Restoration: each individual stem run through iZotope RX 11 for cleanup - de-noise, de-click, spectral repair - since a raw AI-separated stem always carries artifacts a human ear will catch

Second targeted pass: a stem-specific UVR model re-applied where the first pass left bleed or artifacts (vocal isolation models on vocal stems, karaoke models where instrumental bleed persisted)

Musical analysis: confirmed BPM (104) and key (C Major / A minor) directly from the audio, along with the correct chord progression, so the client has accurate reference data for any further arrangement or remix work

Final delivery: clean multi-stem set imported into an Ableton session, organized and ready for production

Off-the-shelf stem separators get you 70% of the way there and stop. The gap between "technically separated" and "usable in a real mix" is restoration work - knowing which artifacts matter, which model to re-run where, and validating the musical fundamentals (key/tempo/progression) so the client isn't guessing at the material they're building on.

Delivered a clean, tempo- and key-confirmed multi-stem set, restored and organized in Ableton - ready for the client to produce further without fighting separation artifacts.