Skip to content

8D-SonKu V2: approved square orbit + lightweight musical FDN and headphone A/B - #629

Closed
masarray wants to merge 15 commits into
feat/8d-sonku-p0-editor-recovery-20261113from
feat/8d-sonku-v2-externalization-acoustics-20261113
Closed

masarray wants to merge 15 commits into
feat/8d-sonku-p0-editor-recovery-20261113from
feat/8d-sonku-v2-externalization-acoustics-20261113

Conversation

@masarray

@masarray masarray commented Nov 13, 2026 •

Copy link
Copy Markdown
Owner

V2 externalization — experimental, reversible, based on the advanced PR #606 GUI/DSP

A user reported 8/8 correct blind directions and 4/4 known, but FRONT/REAR sound close to the head and 360° orbit depth is unnatural. This branch starts at PR #606's verified advanced head rather than overwriting current mixer main.

Implemented

  • Fixed-memory first-order image-source room acoustics: two emitters × four walls, geometry updates every 64 frames, fractional delay with ITD/ILD cue and sample-smoothed gains. Single native JUCE PCM path; no new threads, heap allocation, ray tracing or GUI DSP.
  • Default amount=0: verified original MIT KEMAR, 8/8 correct listener baseline remains unchanged. No native GUI/lease/session change until qualification.
  • Renderer produces original four directions and orbit, plus same-input V2 room versions; two C++ PCM tests assert causality, bounded finite output, direction and buffer partition parity.
  • Research-backed architectural decision with staged M2 ITD-aligned interpolation and M3 reflection improvements. Based on Allen & Berkley image-source method, Steam Audio architecture, 3D Tune-In and externalization literature; no copied vendor source.

Explicit gates

  • Chromium/portable Linux C++ qualification and downloadable WAV A/B; compare perceived front/rear distance and motion continuity at matched perceived volume.
  • Windows exact SHA build, performance benchmark and subjective headphones needed before shipping or enabling the ROOM effect as default.
  • This is stacked on PR 8D-SonKu: native binaural DSP, studio orbit GUI, smooth motion and headphone FAT #606; do not merge it independently ahead of its base. Prevent duplicate Windows FAT runs.

CI and listening qualification — 2026-11-13

Source truth: exact SHA 272b1981c879364c843ed189a7101c3c45fd01b0; stacked on the latest SonKu GUI/core branch PR #606. Compared with PR #606 base: 5 commits ahead, 0 behind, 11 scoped files; no unrelated UI/sends/fader rewrites.

Release guardrails: Stage M1 remains OFF by default on native JUCE so the user's verified 8/8 directions do not regress before front/back A/B. There is no native GUI SPACE control or state schema change yet; no psychoacoustic claim of improved externalization until owner listens to V1 vs V2 at perceived matched level. This is first-order room imagery with approximated reflected ITD/ILD, not true HRIR-convolved early reflections, phase-aligned direct HRIR, personalized HRTF, or FDN late reverb.

Next step: collect listener FRONT/REAR depth and orbit comparison, then decide whether to ship V2 ROOM through typed native ACK/readback with a compact SPACE control, and separately implement/verify proper time-aligned HRIR interpolation. Test exact Windows installer/ASIO once stacked PR #606 is integrated; keep public release gated.

2026-11-13 — double-orbit listener qualification and square-room hypothesis

User feedback: source directions were identified correctly, but RIGHT→FRONT was more convincing than other segments, with perceptually elliptical depth (one lateral side seems farther than FRONT). No claim that the cause is proven. Existing measured KEMAR direct-HRTF model has fixed radius; first-order rectangular room (6.4×8m) causes asymmetric front/lateral reflection distances, while nonpersonalized HRIR interpolation/ITD remains another possible factor.

Delivered:

  • Two continuous 360° cycles in 20 seconds per WAV, 10s/revolution, with no mute/fade/reset at the 10s junction. C++ orbital phase closure and PCM seam energy/format gates.
  • Clockwise and counterclockwise runs for each of 3 conditions: V1 direct measured HRTF, V2 rectangle 6.4×8m, V2 symmetric square 7.2×7.2m, all using the exact deterministic source noise and same direct measured HRIR.
  • Counterclockwise specifically exposes RIGHT→FRONT (7.5–10s / 17.5–20s), clockwise exposes FRONT→RIGHT (0–2.5s / 10–12.5s). Direct audio engine angle signs are unchanged.
  • Square-room diagnostic uses equal 0.70 reflectivity on four walls, no new GUI control / session schema; the default native plugin room processing remains OFF. Test verifies left/right ear mirror symmetry for the candidate.
  • Reconciled the latest parent PR 8D-SonKu: native binaural DSP, studio orbit GUI, smooth motion and headphone FAT #606 changes through merge e1ac0049ef50c2253307212914653443417a12c6, retaining its newer bridge/UI tests/packager.

Evidence: scoped Fast UI/C++20 and Chromium GUI tests, native 44.1/48k FIR/room PCM fixtures, 20s WAV contract and benchmark PASS at exact SHA 3a74b0273c7a7e13d5d2614183d19995eb12ad57. CI run and DOWNLOAD headphones A/B ZIP (artifact ~23.8 MB, GitHub authentication may be required). The six 2x orbit tracks are baseline/rectangle/square for each of CW/CCW; original static and blind trials also remain. Editor screenshots.

CPU from same runner, informative only: 48k/128 V1 48.482ms vs V2 55.354ms per 3s rendered (ratio 1.142; ~1.8% of a CI core); no physical Windows ASIO benchmark. This is an offline research experiment, not evidence of successful human front/rear externalization or proof of a perfect circular sound field. Do not merge or enable room by default without listening qualification and final Windows FAT.

2026-11-13 — listener-approved Square checkpoint + lightweight musical late reverb

Listener result (reported, not lab measurement): the symmetric 7.2×7.2m Square Orbit is now heard as circular and enjoyable. Pinned the exact accepted baseline at checkpoint branch = 3a74b0273c7a7e13d5d2614183d19995eb12ad57. DO NOT alter square ER/HRTF angles while iterating ambience.

Before M2: first-order image-source reflections existed (2 sources × 4 walls) and Square Orbit had been user-auditioned. There was no late reverb and the native insert still defaulted room OFF (experimental).

Delivered M2 on this stack:

  • New original engine/builtin_fx/sonku/SonKuMusicalReverb.h: fixed eight-line orthogonal Hadamard FDN with prepared per-line RT60 feedback coefficients, damping and 12ms wet transition. No dynamic callback allocation, locks, IO, FFT or extra device callback; 8×4096 float fixed rings (~128KiB per SonKu instance), zero DSP work while disabled. Its wet-only tail leaves the KEMAR direct and approved Square ER unchanged.
  • Studio Glue RT60≈0.78s, damping cutoff 5.9kHz; Airy Chamber RT60≈1.28s, 7kHz. Modest A/B wet level 0.42. Both OFF in live/native GUI pending headphone listening. No contract/session schema/GUI feature yet.
  • Standalone finite/impulse/causality/stereo-tail/decay/preset/bypass C++20 tests; native core active-room buffer partition parity; same-runner CPU benchmark at 44.1/48k and blocks 128/512.
  • Two new 20-second 2×360° broadband A/B WAVs on identical source compared against approved sonku-v2-square-orbit-cw-2x.wav, plus three 20-second original synthesized plucked-tone tracks to assess metallic artifacts and musical pleasantness with baseline/Glue/Airy.
  • Exact SHA 880e158e1cbbabb9ce0ed13bedd5f3240a502566 PASS fast C++/contract tests, genuine Chromium GUI assembly/editor ACK mock test, native C++ offline renderer, reference/blind/2-turn WAV contracts and headroom checks: GitHub Actions run.
  • Download user A/B ZIP: https://github.com/masarray/arsonmix/actions/runs/43364802118/artifacts/14010185867 (GitHub login may be required); contains WAVs, read LISTENING-GUIDE.txt. Editor screenshots.
  • Cost (verified on the latest exact-SHA runner): 48k/128, V1 57.520ms / V2 ER 66.072ms / V3 ER+FDN 72.324ms to render three seconds; incremental FDN +9.5% vs V2, V3 ~2.4% of one Linux CI core for one instance. This is an informative benchmark only, not Windows ASIO/hardware latency or a guarantee for users. See run 43364802118.
  • Research: original implementation guided by Stautner/Puckette orthogonal delay networks, Jot/Chaigne RT60 analysis, and Valhalla guidance on avoiding metallic short reverbs; no proprietary code or commercial GUI copied.

No premature release: this branch remains a draft stack over #606, not main/public installer. Square geometry is approved perceptually by the owner, FDN musicality not yet approved. Need level-matched owner listening on real headphones, exact-SHA Windows native/ASIO qualification, and typed GUI/lease/session design before making ROOM/FDN controllable in the released plugin. Preserve A/B bypass as reversible control.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 13, 2026 •

Copy link
Copy Markdown

Deploying arsonmix with  Cloudflare Pages  Cloudflare Pages

Latest commit: 63a85fb
Status:⚡️  Build in progress...

View logs

@masarray masarray changed the title 8D-SonKu V2: bounded image-source room A/B, PCM regression and headphone qualification 8D-SonKu V2 — measured early reflections, headphone A/B and realtime CPU benchmark Nov 13, 2026
@masarray masarray changed the title 8D-SonKu V2 — measured early reflections, headphone A/B and realtime CPU benchmark 8D-SonKu V2: dual-rotation orbit, square vs rectangular room A/B and realtime qualification Nov 13, 2026
@masarray masarray changed the title 8D-SonKu V2: dual-rotation orbit, square vs rectangular room A/B and realtime qualification 8D-SonKu V2: approved square orbit + lightweight musical FDN and headphone A/B Nov 13, 2026

Copy link
Copy Markdown
Owner Author

Audit checkpoint — 2026-11-13 (live repo, no premature release)

Source inspected: main 7232c1ebe9e4; parent #606 76271c5c274f; this stacked #629 63a85fbb010d (after audit test-only fix).

Verified complete on experimental stack

  • Approved 7.2 m square / fixed HRIR + 2 emitters × 4 early-wall paths; checkpoint 3a74b0273c7a remains pinned and untouched.
  • An 8-delay fixed-memory Hadamard FDN exists in SonKuMusicalReverb.h, rendered inside the single SonKuSpatialCore. Studio Glue (~0.78 s) / Airy Chamber (~1.28 s) are offline-only candidates.
  • Prior exact-SHA 880e158e... fast DSP/PCM, SonKu Chromium FAT and WAV publishing PASS (run 43364802118); not a physical ASIO listening qualification.
  • Audited and repaired regression test: 63a85fbb010d now compares Square+FDN against an identical Square with FDN disabled, verifying exact initial PCM identity and late-only difference at 44.1/48 kHz with partition-parity checks. Before this fix the comparison unintentionally used Rectangular+OFF, so it could not isolate the FDN.

Not done / release gates

  1. 8D-SonKu: native binaural DSP, studio orbit GUI, smooth motion and headphone FAT #606 and 8D-SonKu V2: approved square orbit + lightweight musical FDN and headphone A/B #629 are still draft. 8D-SonKu: native binaural DSP, studio orbit GUI, smooth motion and headphone FAT #606 diverges from main by 7 newer commits touching wire GUI / updater / Main.cpp; reconcile main safely into 8D-SonKu: native binaural DSP, studio orbit GUI, smooth motion and headphone FAT #606 first, preserving V61–V64 WIRE and updater semantics, then forward-port 8D-SonKu V2: approved square orbit + lightweight musical FDN and headphone A/B #629. Do NOT merge the stack blind or overwrite the accepted square branch.
  2. SonKuSpatialProcessor presently queues/persists only sonku::Settings (azimuth, orbit speed, spread, spatial mix, trim, moving, clockwise). There is no native typed Room/FDN intent, PCM-applied ACK/readback, UI control, or versioned .smix state migration. OFF is still the live default; Square+FDN WAV does NOT imply available native ambience.
  3. Reverb setters are OFFLINE/CONTROL-side only. setAmount() currently can clear 8×4096 floats, and setPreset() computes exponential coefficients. They must NOT be called directly from the realtime callback when future GUI control is added. Prepare immutable/coefficient state outside callback, pass bounded POD/SPSC intent with one DSP authority, apply with smoothed level, no RT allocation/blocking/big clearing and only report APPLIED after PCM readback.
  4. Native getTailLengthSeconds()==0 presently describes the no-FDN product. Reassess when enabling wet tail so future host/graph management doesn't truncate reverb; test bypass, mute, FX reorder, remove, device-restart and tail reset.
  5. Support policy is currently 44.1/48 kHz only. Qualify stereo-only behavior, unsupported rates, device/reconnect, multi-instance memory/CPU, safe session restore, and Windows ASIO p95/p99/xruns before any public installer.
  6. Subjective musicality is still unapproved. Listener should compare same-input Square baseline vs Glue vs Airy with real headphones at matched perceived level; reject/flutter/metallic/tail-masking cases before adding compact SPACE/AMBIENCE UI. The earlier Square Orbit preference remains the authoritative acoustic checkpoint.

Low-cost CI: run cheap C++/PCM + native ABI/editor/contract checks on the next focused change; reserve one exact-final-SHA Windows FAT/ASIO pass after reconciliation, not redundant full Windows jobs on every iteration. Do not confuse green CI for hardware proof.

Copy link
Copy Markdown
Owner Author

Integrated SonKu consolidation checkpoint — 2026-11-13

The acoustic/native stack is now reconciled into PR #649, an integrated draft based on current main. Its two-parent merge carries this PR's history and the latest #606 native/GUI/updater/WIRE parent simultaneously. Please do not merge this older stacked PR independently into main, or queue redundant Windows builds: the canonical combined qualification now belongs to #649, after which old drafts can be closed as superseded.

Preserved: approved Square Orbit checkpoint 3a74b0273c7a7e13d5d2614183d19995eb12ad57, 8-line FDN, fixed ring/no-RT-clear behavior, source-distance + room models, typed native SPACE controls, schema v2 session restore, real Chromium ACK/readback and musical A/B WAV. All remain pending actual Windows/ASIO and listener release gates.

Copy link
Copy Markdown
Owner Author

MILESTONE A: historical research PR archived as CLOSED UNMERGED. This original early SonKu HRTF, room, distance or FX-insert experiment was superseded as a release path by owner-approved P9.9 #733. Not all source was necessarily imported; the experimental branch remains available for exact code recovery. Original source branch feat/8d-sonku-v2-externalization-acoustics-20261113, head SHA 63a85fbb010d6e5590ed7bfe1673eb5bed50af1f, target feat/8d-sonku-p0-editor-recovery-20261113. Full disposition and immutable heads: docs/governance/MILESTONE_A_PR_BRANCH_RECONCILIATION_20261117.md on chore/milestone-a-pr-history-20261117. No branch/checkpoint deletion, no merge, no DSP/GUI change, no release. Reopen only under a separately scoped task after approval.

@masarray masarray closed this Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant