BELIEVEYOUME stems vs the stem player

What is actually on disk for Shiner's BELIEVEYOUME, and what the live stem player exposes of it.

BUILT 2026-08-17 · SOURCE: /Volumes/T7 Shield/Moises_Stems/BelieveYouMeStems/ · SITE: drwu-htmls.vercel.app/believeyoume-stems

1. The short answer

2. Per-song stem census (what is on disk)

Root: /Volumes/T7 Shield/Moises_Stems/BelieveYouMeStems/, the complete root, holding all 10 album tracks. Counted from the filesystem, macOS ._ resource forks excluded.

Highlighted rows are the eight multi-guitar songs. The Voc column never exceeds 1. That is the measured result.

The vocal question, settled

I swept five roots for any file whose name contains vocal, vox, backing, harmon, lead voc, choir, double, whisper, or scream:

• T7 Shield/Moises_Stems/BelieveYouMeStems/: 10 hits, all named exactly vocals, one per song
• T7 Shield/Stems/BELIEVEYOUME/: 0 hits
• T7 Shield/BandBooks/BELIEVEYOUME/: 2 hits, both vocals.wav from a demucs 6-stem re-separation
• T7 Shield/Music to Stem/Beleiveyoume/: 0 hits (source FLACs, un-stemmed)
• Moises_Input/stems/htdemucs_6s/: 0 hits

No lead/backing split, no harmony stem, no doubles exist on disk today. Getting a multi-vocal player means re-separating with a model that splits vocals. That is a new render. No unused file is sitting there waiting.

3. What the player exposes

Source: drwu-htmls.vercel.app/believeyoume-stems, audio from drwu-stems.vercel.app/believeyoume-charts/stems/.

The page hard-codes six lanes for every song, with no per-song lane list:

const LANES = [
  {key:"drums",  label:"Drums"},
  {key:"bass",   label:"Bass"},
  {key:"guitar", label:"Guitar"},   <-- one lane; 22 disk stems fold into 10 of these
  {key:"piano",  label:"Keys"},
  {key:"vocals", label:"Vocals"},
  {key:"midi",   label:"Guide"}
];

I probed all 12 songs × 6 lanes on the live host with ranged GETs. Every lane returns 206 with content-type: audio/mpeg; a deliberate bogus path returns 404, so the probe distinguishes real files from a blanket fallback. All 70 expected lane files are present and real. The two midi gaps are declared in the page's own noLanes field, so they are honest.

4. The contrast, measured

4a. The guitar lane is a rhythm-plus-solo sum, and the solo is buried

measured I decoded the site's Mutiny lanes and envelope cross-correlated them against every Mutiny guitar stem on disk, with lag search. Control first: the site vocals lane against the disk vocals stem must top the table, or the measure is worthless.

measured Levels explain the burial. On disk, Mutiny Rhythm Guitars peaks at −0.0 dBFS (mean −17.3), Solo Guitars peaks at −2.9 dBFS (mean −28.4). Summed into one lane, the solo is about 11 dB under the rhythm. Pressing the SOLO button on the player's guitar lane cannot separate them, because they were summed before encoding.

4b. What each song loses at the lane boundary

Jackie and Asleep in the Trunk get a Guitar and a Keys lane despite neither folder having one. A demucs 6-stem re-separation of exactly those two songs is at BandBooks/BELIEVEYOUME/3-guitar-derivation/_demucs_6s/htdemucs_6s/ with guitar.wav and piano.wav, which accounts for it.

4c. Fidelity

measured Disk stems are 44,100 Hz stereo WAV, roughly 44 MB per Mutiny stem. The site serves 96,021 bps CBR MP3, about 3.0 MB per lane. Equal-duration lanes therefore produce byte-identical file sizes. To rule out duplication I md5'd all six Mutiny lanes, and every one differs, so the lanes carry genuinely different audio. The house bar of 48 kHz / 24-bit applies to sample delivery, so 96 kbps counts as a deliberate streaming choice here.

4d. Two songs on the player are not on the album

measured The player's dropdown lists 12 songs. The album source folder Music to Stem/Beleiveyoume/BELIEVEYOUME (2025)/ holds exactly 10 FLACs. Number One Crush and Only Happy When It Rains are on the player, serve real audio for 5 lanes each, and have no stem folder in any BELIEVEYOUME root. inference Both titles are Garbage songs, so a covers session or a cross-project spill would explain it, and I have not confirmed which.

5. Four traps in the stem folders

6. Method and provenance

• measured File census: find over five roots, ._ resource forks excluded, roles parsed from filenames.
• measured Levels: ffmpeg volumedetect, whole-file mean and peak, plus band splits at <120 Hz, 250 Hz–2.5 kHz, >6 kHz.
• measured Source attribution: 20 ms RMS log-envelope, FFT cross-correlation with ±10 s lag search. Validated before use. A plain zero-lag sample correlation put the vocals control at r=−0.02 and was discarded as unusable; the envelope method put the same control at r=+0.847 as top match.
• measured Live host: ranged GETs reading status, content-type and content-range; a bogus path returned 404 to prove the probe separates real files from a fallback.
• inference Only three claims are inference and are tagged as such: the account for the two non-album songs, the origin of the l label, and the reading that g3 was meant as a guitar slot.
• Per the stem-detection gate, a pitch found in one stem does not prove that instrument played it. Nothing here attributes an instrument to a part; this page reports which files exist, how loud they are, and which of them the player is serving.

If you want the multi-guitar detail on the player

The stems that would carry it exist today for 8 songs and need no new separation. The player needs a per-song lane list in place of the hard-coded six-lane array, since the lane sets genuinely differ (Mutiny has an L/R pair nothing else has, Not Too Much has strings, three songs have only lead/rhythm). Multi-vocal needs a re-separation first, because no file on disk supports it.