Apollo texture evaluation — real music and memory evidence

Review material for JusperLee#44 — not an approved Apollo release.

Eight aligned 15-second real-music excerpts with raw WAVs (source proxy, degraded input, upstream Apollo at 0 dB and that exact prediction at 1/2/3 dB), randomized BS.1770 loudness-matched blind copies, a single-player HTML audition page, source hashes, reproduction scripts and memory measurements. No lossless clean master or controlled listening result is claimed. The low-level case has a documented -18 dB input gain before inference. All cases, including objective adverse/caution results, are retained.

The full-output STFT remains expensive: a 180-second probe repeating real Apollo output measured about 2328.6 MiB peak tracked allocations / 3042.2 MiB process RSS, excluding the model.

33 tests passed locally and on pinned Linux CPU CI: https://github.com/hemanpapa-coder/Apollo/actions/runs/38918654428 . Upstream CI requires maintainer approval. Keep PR Draft pending listening.

Audio attribution (full credits and changes in bundle README):

• Where You Are Now — Benjamin Orth featuring Wolf Fly; original vocal by Ascension / Rhiannon, CC BY 3.0: https://ccmixter.org/files/Benjamin_Orth/33036
• Studies for Piano — Richard Kearns, CC BY 3.0: https://commons.wikimedia.org/wiki/File:Study2.ogg
• Vibe Ace — Kevin MacLeod, CC BY 4.0: https://librosa.org/data/audio/Kevin_MacLeod_-_Vibe_Ace.toml
• Dance of the Sugar Plum Fairy — P.I. Tchaikovsky (composer), Kevin MacLeod (arranger), CC BY 4.0: https://librosa.org/data/audio/Kevin_MacLeod_-_P_I_Tchaikovsky_Dance_of_the_Sugar_Plum_Fairy.toml

https://creativecommons.org/licenses/by/3.0/ and https://creativecommons.org/licenses/by/4.0/ . Audio was cropped, optionally re-encoded, restored and post-processed by Neuracoust for evaluation; no artist endorsement is implied. Source recordings remain under their own licenses; code retains upstream CC BY-SA 4.0.

ZIP SHA-256: 528da89b7cb2bc5b18463f26bfa47229ed90d67ecfc50096ef0f89080dcb44d1
Size: 621114458 bytes. Local full ZIP CRC validation passed (142 files, 120 WAVs including context).

Suno Untitled1017 — Apollo comparison data (2026-10-17)

Full-song application comparison completed (2026-10-19)

Download the complete application comparison ZIP · Release and provenance

All 25 windows of the creator-authorized 218.36-second Suno song are complete. Extract the ZIP and open enhanced-full/index.html: one player switches at the same position between resampled input, upstream Apollo, and the full application chain. Three aligned tracks are also supplied in comparison.rpp (generated session; live DAW playback untested), alongside raw enhancement, hashes, exact local runners and numerical reports. All listening versions are 44.1 kHz stereo and matched to -23 LUFS (spread <0.000001 LU); 4x-oversampled peaks are -10.527/-9.850/-9.919 dBFS respectively.

The full chain uses unchanged official weights with polarity/gain/shift ensembles, fusion, vocal/phase stabilization and 2 dB texture softening. This supplements, rather than changes, the narrow default-OFF CLI option. A real 14-second official-checkpoint prediction from the disk-stage runner matched its saved reference exactly in float32 (max absolute error 0); interrupted-resume and corrupt-checkpoint tests passed. This is numerical evidence for that input, not a full-song or cross-device bit-identity claim.

Resource record: first 13 windows MPS, later windows CPU; resumed CPU uses 3 threads, 35% execution duty, whole-system CPU/RSS/swap backoff and atomic disk stages. Latest resumed inference peak RSS 823.6 MiB; separate postprocessing peak 474.1 MiB. Limits are sampled guards, not instantaneous caps. Resumed inference wall time 3712.88 s excludes earlier attempts and pauses and is not a fair speed comparison with the upstream run. The disk policy does not change model weights or audio context.

One song, no clean reference or controlled listening verdict. Existing CI and earlier adverse cases remain documented below; upstream Actions still require maintainer approval. Keeping Draft for listening and maintainer review.

---

Full-song ZIP SHA-256: 0dddee5949adb5ca88c5254367109d7fe704b82020575b6d4877193952aea030 (285822013 bytes).

Scope clarification and full-method source

The original large Suno comparison ZIP below contains texture-only 0/1/2/3 dB processing, not the full application enhancement. The full Neuracoust method also uses polarity ensembles, a gain-compensated +6 dB prediction, eight shifted-context ensembles, phase-preserving fusion, vocal-band consensus and phase stabilization.

The added Apollo-full-enhancement-method-20261017.zip contains the four DSP sources, a fixed-checkpoint reproduction wrapper, license and 19 passing local tests. It is also reviewable in PR #44. Source ZIP SHA-256: 6ae67673d6d1c237838f102d8c782065aa6ba05899b4681e94514a93e639bc05 (16990 bytes). Completed full-song audio is linked above. The CLI integration in PR JusperLee#44 remains the default-OFF final texture stage.

---

Evaluation data for Apollo PR #44, not a production software release.

The creator supplied Untitled1017.wav, states that they created it with Suno, and explicitly authorized public GitHub distribution of these comparison files on 2026-10-17. No additional blanket audio license is asserted.

Contents: original 218.36-second 48 kHz stereo WAV; full-length 44.1 kHz input and Apollo 0/1/2/3 dB outputs; eight aligned 15-second excerpts, raw and randomized loudness-matched A–E copies; HTML audition page; five-track REAPER comparison session; measurements, SHA-256 manifest, reproduction script and instructions. Extract the ZIP and open index.html. REAPER session structure was generated; live DAW playback is untested.

No added codec degradation. Input resampling uses scipy resample_poly (147/160). The full track was inferred once with upstream e84bcac, official checkpoint SHA-256 99d9af7f1ff20e63c393035513a655392818d66b4d7fc23d658175c1f15e8d76, Torch 2.8.0 / MPS, chunk/overlap/pad/batch 6s/1s/1s/1. The exact same prediction was processed at 1/2/3 dB with contribution code 4a89524.

All eight blind groups passed read-back checks: maximum within-group loudness spread 0.000000631 LU; maximum 4x-oversampled peak -9.059 dBFS. Inference took 113.07 seconds. These timing/peak checks do not establish audible improvement.

These are evenly spaced excerpts from ONE song, not eight independent songs or listener-confirmed instrument/adverse categories. No clean reference or controlled listening verdict. This supplements the earlier CC-recording dataset. Existing 33-test CI covers unchanged processor code; this dataset is separate rendering validation. Keep PR Draft for listening and maintainer review.

ZIP: 789188701 bytes. SHA-256: 90592aedd3b599ea94a862b17bf28c05456f603bb29df57a053d35eb7419945b. Local ZIP CRC verification passed; all 86 WAV hashes matched the saved manifest.