title: How Much Information Does Stem Separation Actually Lose?
subtitle: Reconstruction fidelity across 2-stem, 4-stem, and 6-stem modes on 50 MUSDB18-7s
  tracks
run_date: March 14, 2026
run_timestamp: '2026-03-14T11:09:58'
dataset: MUSDB18-7s
dataset_url: https://zenodo.org/record/1117372
n_tracks: 50
key_finding: Splitting into 6 stems vs 4 stems increases reconstruction error by 1.2
  dB (-21.6 dB for 4-stem vs -22.8 dB for 6-stem relative to original)
methodology_note: Each mixture was separated into stems, which were then summed back.
  The difference between the reconstruction and the original mix measures information
  loss. Lower MSE, higher correlation, and lower dB difference all indicate better
  reconstruction fidelity.
modes:
- stems: 2
  model: htdemucs_ft
  label: 2-stem (vocals + accompaniment)
  n_tracks: 50
  mean_mse: 0.000169
  mean_correlation: 0.9961
  mean_db_difference: -21.7
- stems: 4
  model: htdemucs_ft
  label: 4-stem (vocals, drums, bass, other)
  n_tracks: 50
  mean_mse: 0.000169
  mean_correlation: 0.996
  mean_db_difference: -21.6
- stems: 6
  model: htdemucs_6s
  label: 6-stem (+ guitar, piano)
  n_tracks: 50
  mean_mse: 0.000138
  mean_correlation: 0.9967
  mean_db_difference: -22.8