Shiner, BELIEVEYOUME (2025), track 08. Splitting the single Moises
vocals stem into a centered lead and the panned stack around it.
Track 08 did get the fullest Moises treatment on the record: six kit parts, three guitar splits,
keys, bass, other, other-drum, and a strings stem that no other BELIEVEYOUME song has.
The vocal came out of that pass as one combined file.
So the split below was made here, from that combined stem.
The obvious suspect for a hidden vocal was strings, since track 08 is the only song
that has one. Measured over the loudest vocal minute, it is not a voice.
| Stem | dBFS | Sibilance 4-10 kHz | Envelope corr with vocal | Verdict |
|---|---|---|---|---|
| vocals | −19.5 | 0.3153 | 1.000 | the only voice |
| strings | −32.1 | 0.1227 | −0.109 | pad, sounds when the vocal rests |
| keys | −27.6 | 0.1127 | +0.386 | follows the vocal, no consonants |
| Solo Guitars | −24.9 | 0.0917 | −0.286 | guitar |
| Other Guitar | −43.8 | 0.1054 | −0.006 | guitar |
strings measures with the guitars, and it runs
anti-correlated with the lead, meaning it sounds when the singing stops. Moises also labels its
outputs literally (Rhythm Guitars, Hi Hat, other drum), so a
vocal split would have arrived named Lead Vocals and Backing Vocals.
Neither filename exists on any disk.
Per-bin azimuth masking on a 4096-point STFT. For every time-frequency bin, two things get measured: how evenly the energy divides between left and right, and whether the two channels are in phase. Bins that are both balanced and coherent are the centered lead. Everything else is the stack.
panbal = 1 - |(|L| - |R|)| / (|L| + |R|) 1 = dead centre, 0 = hard panned coh = (cos(∠L - ∠R) + 1) / 2 1 = in phase mask = (panbal * coh) ^ 2.5 lead = L*mask , R*mask backing = original - lead exact complement
Defining backing as the complement means the two files sum back to the source sample for sample, so nothing is invented and nothing is lost.
| Measure | Original | LEAD | BACKING | Reading |
|---|---|---|---|---|
| L/R correlation | 0.697 | 0.945 | 0.274 | lead centred, stack wide |
| Median f0 | 192.5 Hz | 194.7 Hz | 180.6 Hz | stack runs lower |
| f0 spread (IQR) | 37.4 | 40.1 | 44.7 | stack more scattered, harmony-like |
| Sibilance 4-10 kHz | 0.3153 | 0.3068 | 0.3267 | stack is voice, not reverb |
| Voiced frames | 65.7% | 64.9% | 58.7% | both carry real singing |
| Reconstruction error | 2.78e-17 max abs | lossless complement | ||
Brandon supplied a genuine unlabeled Moises Premium backing-vocal export for Endless Summer. Measured against that song's lead stem it scores an 88.3% consonant-interval share over 508 co-voiced frames, against a 40.0% flat baseline, peaking on unison 34.4% and minor third 29.2%. That gives a real number for what a correct lead/backing pair looks like, so the three methods below are scored against it rather than against my own description.
| Method | Lead L/R | Backing L/R | Sib ratio | Env corr | Harmony % | Verdict |
|---|---|---|---|---|---|---|
| DSP azimuth | 0.928 | 0.138 | 0.96 | +0.685 | 93.8 | PASS |
| MDX on vocal stem | 0.851 | 0.422 | 0.69 | +0.009 | 78.8 | FAIL |
| MDX-guided | 0.806 | 0.372 | 0.78 | +0.405 | 90.2 | FAIL |
UVR_MDXNET_KARA_2) lost twice. Fed the isolated vocal stem it returned a backing that
barely correlates with the lead at +0.009 and carries only 0.69 of its sibilance, which is residue and
reverb rather than a second voice. That is the out-of-distribution failure predicted before the run,
since karaoke models are trained on full mixes. Fed the full mix and used as a spectral guide it
improved to 90.2% harmony and still missed the sibilance gate at 0.78.
~/tools/MSS-Training has configs but zero weights),
so it needs a multi-gigabyte install first./Volumes/T7 Shield/Moises_Stems/BelieveYouMeStems/08 - Not Too Much/.