The Quiet Room - Guide Track Render Diagnostic Summary
The Quiet Room - Guide Track Render Diagnostic Summary
Received: 2026-07-30 10:59 GMT+1 via Telegram from John Source: External thread context document covering Songs 1-6 + guide track v1/v2 diagnostics Track: The Quiet Room v6g · 65 BPM · A minor Files saved:
- `diagnostics/thread-context-full-analysis.md` - full thread history
- `diagnostics/render-diagnostic-summary.md` - this file
Core Diagnosis: AUDITORY FUSION, not masking
Dakota's root-cause hypothesis ("MuseScore piano patch masks everything") is wrong. Refuted three ways:
1. Differential audibility: bass +21.9 to +29.6 dB, strings +7.3 to +19.0 dB, harmonics +6.5 dB - all plainly present 2. Choruses are STRING-dominated: 57-60% of chorus energy sits in 420-700 Hz (strings' register). Piano fundamentals are 175-440 Hz 3. Section arc EXISTS: verse -31.6 dB → chorus -22.4 dB = +9.2 dB lift, 17.5 dB range, LRA 14.2 LU
Why it sounds like one instrument:
All four source-separation cues are defeated simultaneously:
- **Onset**: In C1, piano+strings+bass+pad all start on beat 0
- **Envelope**: All share one struck-decay envelope (nothing sustains - drops -33 to -40 dB by 1.2s)
- **Pan**: 45/50/55/60/64 on 0-127 scale = all within ±15% of centre (effectively mono)
- **Timbre**: Programs 48 (String Ensemble 1) and 49 (String Ensemble 2) are near-identical patches
Five Concrete Bugs
1. Outro never rendered: duration_beats:144 but form ends at beat 152 2. No chorus escalation: C1 -22.54, C2 -22.33, FC -22.35 = 0.21 dB spread. Strings+harmonics pinned at velocity 127 3. -24.1 LUFS = ~10 dB too quiet for upload 4. Pad inaudible: A3=220Hz in UNISON with piano A3. Duplicate pitches in chords [A3,E3,A3] and [F3,C3,F3] 5. Harmonics: E6 +2.4 dB and C6 +2.9 dB below detectability; only A6 clears
JSON Annotations vs Reality
The pre-render gate validates annotations, not the music. Multiple annotations are false:
- `v3_to_fc_jump`: claims guitar 35→90 (+157%); actual 30→35 (+17%)
- `velocity_drop_from_c2`: claims 78→50, 75→50; those velocities appear nowhere
- Bridge caesura: claims beats 12-16 total silence; bass runs to beat 15, pad to beat 16
- `high_element_enters`: claims A5 880Hz; note is A6 1760Hz
- `frequency_map` claims 200-500Hz left clear; strum includes C4, F4, G4, A4 in that range
- GT-1 (no vocal in 200-500Hz) violated by the guide's own notes
- `duration_seconds` ~133s but 152 beats at 65 BPM = 140.3s
Sustain Problem (unresolved)
Nothing sustains. All notes decay -33 to -40 dB by 1.2s. Either:
- Program changes aren't reaching FluidSynth (everything renders as piano), or
- Presets load without sustain
Need the .mid file or five single-channel renders to settle it.
Ranked Fixes
1. Stagger onsets - add `onset_offset_ms` per part so instruments don't all hit simultaneously 2. Spread pan - 18-110 range, not 45-64 cluster 3. Differentiate envelopes - vary attack/sustain/release per instrument 4. Stop using velocity 127 - strings and harmonics pinned at 127 leaves no headroom 5. Restore piano - ~52/72/40/84 range, not 25-35 6. Compute duration_beats correctly - 152 not 144 7. Gain +9dB post-render - target -14 to -15 LUFS 8. Fix pad voicing - [E3,A3,C4] not [A3,E3,A3] (avoids unison with piano) 9. Make Bridge caesura real - actually silence other parts 10. JSON format overhaul - add `onset_offset_ms`, `reverb_send`, `attack`, `target_lufs`
Cross-Song Context
This is song 6 in an iterative analysis thread training Dakota as a local AI composer writing dark folk for Suno. Key persistent failures across all songs:
- Vocal buried in choruses (never fixed in 6/6 versions)
- No energy above 500 Hz (all 4 versions)
- Hook motif contour self-similarity near zero (both A and B)
- Both versions deflate final chorus
Three-song trend (Version A):
- Chord vocab: 17→22→15 (fixed)
- LRA: 7.1→6.5→10.1 (big gain)
- Chorus ceiling lift: +3.4→-0.7→+6.2 st (best yet)
- Hook r: -0.080→+0.480→+0.165 (REGRESSED)
- V1→Ch1 vocal swing: -7.4→-7.4→-7.7 dB (NEVER FIXED)
Guide v2 key conclusion:
Suno faithfully reproduces a signal whose acoustic signature IS background humming. 13 metrics at 1.9x to 206x deficit vs a real lead vocal. Top recommendation from analysis: sing/hum the guide into a phone. Fixes all 16 metrics at once.
Action Items
- [ ] Revise guide-track.json to v7 with de-fusion parameters:
- [ ] Investigate sustain problem - render 5 single-channel passes to check if program changes reach FluidSynth
- [ ] Fix pre-render gate to validate note data, not just annotations
- [ ] Re-render guide track v7
- [ ] Send new render to John for perceptibility check (GT_3)
- [ ] `onset_offset_ms` per instrument (stagger entries by 20-80ms) - [ ] Pan spread across 18-110 (not 45-64) - [ ] Different reverb_send per layer - [ ] Different attack characteristics per instrument - [ ] Remove velocity 127 - use 60-90 range with dynamics - [ ] Fix piano velocity to 52/72/40/84 range - [ ] Fix duration_beats to 152 - [ ] Fix pad voicing to [E3,A3,C4] - [ ] Make Bridge caesura real (silence other parts) - [ ] Add `target_lufs: -14` and post-render normalise