Skip to content

SonKu P9.4 DRAFT — front/back externalization A/B while preserving 3D ellipse and accepted music baseline - #723

Closed
masarray wants to merge 15 commits into
feat/sonku-p93-circular-stage-movable-power-presets-20261116from
research/sonku-p94-front-back-depth-20261116
Closed

masarray wants to merge 15 commits into
feat/sonku-p93-circular-stage-movable-power-presets-20261116from
research/sonku-p94-front-back-depth-20261116

Conversation

@masarray

Copy link
Copy Markdown
Owner

SonKu P9.4 — front/rear perceived-distance A/B (PRIVATE, draft)

Reconciled baselines: operator-approved checkpoint #720 SHA 763706a15a2617a10d23b5e2e43eb5689e454ccb; GUI PR #721 SHA 3648cee51bda236fd7917d04c2a3b40972d3bb17. The 148×83 SVG perspective ellipse is correct and remains unchanged. Neither main nor checkpoint was modified.

Actual user defect: the speaker audio trajectory sounds flattened in the front/back direction, though the visual is a coherent circular orbit viewed obliquely in 3D. A generic KEMAR HRTF can lose front/back externalization on the median plane. This PR is an isolated acoustic hypothesis, not a geometry fix, EQ boost, headtracking claim or dearVR parity.

P9.4 private DSP A/B (requires the ARMIX_SONKU_EXPERIMENTAL_P94_FRONT_BACK_DEPTH CMake option, OFF by default, P9 required):

  • Fixed 361-point cos^4(head-relative centre azimuth) lookup, 45ms smoothing on PCM. 0°/180° = strongest correction; 90°/270° = strict unity. Zero new allocations/threads, no audio FFT, no novel output bus.
  • At FRONT/REAR only when ROOM+EARLY are active: measured direct HRIR level down by up to 0.045 (−0.40 dB), independently processed binaural directional early reflections up to +0.28 (+2.14 dB), a bounded DRR pressure-ratio proxy ≈ −2.54 dB. Preserve all measured pinna kernels, ITD, ILD, original world speaker bass and room modalities, fixed source radius, phase and gain semantics, eight-line late tail. No +7dB output global normalization.
  • Existing mode 9 and persisted .smix schema remain untouched. Only a genuine opt-in native build announces p94FrontBackDepthCue=true; GUI then truthfully displays P9.4 · FRONT/REAR DEPTH (A/B). Older binaries retain P9.1/P9.2 label; no new selectable phantom mode.
  • Native tests sample full stereo PCM at 44.1/48 kHz across 0/90/180/270, room ON/OFF, anti-phase, 64/256 frame exact parity, finite output and 0.98 linked guard; isolated cue checks front/rear equivalence, lateral neutrality, head-yaw and max DRR correction. Previous P9.1 PCM, P9.0, P8.9, GUI and session readback regressions still gated.
  • Exact-source Linux + Windows native JUCE, CTest and portable packaging workflow added to existing scoped experimental FAT path; no installer/release.

CRITICAL: Directional DRR level manipulation is only a hypothesis backed by general externalization research. Numerical values are not externally validated. If it causes audible pumping, muddy/echoey room, front/rear still near-head, bass weakening, tonal shifts or unmatched loudness, REJECT P9.4 and revert to operator-approved checkpoint. Full field listening/ASIO FAT and comparison with dearVR still pending. See docs/SONKU_P94_FRONT_REAR_EXTERNALIZATION_FAT_20261116.md.

Do not merge into main while this PR is draft or call it an acoustic fix until FAT demonstrates benefit.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 16, 2026 •

Copy link
Copy Markdown

Deploying arsonmix with  Cloudflare Pages  Cloudflare Pages

Latest commit: 766b8d7
Status: ✅  Deploy successful!
Preview URL: https://8b988d3f.arsonmix.pages.dev
Branch Preview URL: https://feat-sonku-p95-immersive-cin.arsonmix.pages.dev

View logs

Copy link
Copy Markdown
Owner Author

Evidence update: exact-source Linux PCM A/B passed at 766b8d722ac18354e5c8e6eaa4e81f40ba42bcc8

Actions run #43790492060, job SonKu P9 exact-source C++ PCM preflight: success. Real code compiled as two separate native binaries from exactly the same test source, P9.1+P9.2 reference without P9.4 macro and P9.4 candidate with opt-in macro. The comparator rejects >0.5dB steady-state RMS change and any lateral/ROOM OFF change.

rate 0° FRONT 90° RIGHT 180° REAR 270° LEFT
44.1k −0.3791 dB 0.0000 dB −0.3787 dB 0.0000 dB
48k −0.3826 dB 0.0000 dB −0.3820 dB 0.0000 dB

ROOM OFF comparison is exact, as is P9.1/P9.2 PCM at 90/270 on these fixtures. Calculated bounded direct/early gain ratio at 0/180 is −2.544dB; this is an electrical ratio proxy, not validated acoustic DRR or headphone SPL.

Additional diagnostic finding: with operator-selected LARGE 28.8m room and 1.6m speaker radius, the current first-order horizontal wall images imply approximately 75–84ms excess propagation versus the direct sound. A physically nearer floor image is ~5ms and ceiling ~29ms at the current 1.45m height. Therefore broad late-wall energy boosting is NOT a defensible substitute for a correctly measured early binaural room response. P9.4 remains a deliberately bounded hypothesis; reject if it sounds echoey or only quieter. Do not silently widen the correction during code review; separate low-latency floor/ceiling BRIR research and HRTF/headphone personalization should follow measured headphone evidence.

Windows JUCE+CTest (including native exact-source P9.4 baseline/candidate CTest), portable packaging, acoustic hearing FAT and ASIO device check remain separate gates. GUI perspective ellipse remains unchanged from checkpoint.

Copy link
Copy Markdown
Owner Author

P9.4 exact-source Windows portable research FAT — PASS (2026-11-16)

Head 766b8d722ac18354e5c8e6eaa4e81f40ba42bcc8 completed Actions #43790492060 successfully. Linux source PCM + strict P9.1-vs-P9.4 level comparator PASS; native fast contracts PASS; Chromium UI smoke PASS. Windows JUCE 8.0.15 + WebView2 1.0.3485.44 built the actual mixer GUI, and 7/7 targeted CTest PASS: P8.9 room, P8.9 JUCE state, P9.0 directional externalization, P9.1/P9.2 source, P9.4 source, P9.4 baseline vs candidate executable comparator, and P9 v11 JUCE ACK/legacy restore.

Windows x64 portable artifact #14192672665. Downloaded archive ArSonMix-SonKu-P94-FrontRear-Experimental-Windows-x64-43790492060.zip is CRC-clean, SHA256 2dd10252271769c5432f7ead48e0539c9dd98e39373e4f5fb3a6924a9f4cd529, 8,331,275 bytes. Native mixer exe 15,115,264 bytes. BUILD_PROVENANCE.json exact sourceCommit matches HEAD, experiment mode P9.4_FRONT_REAR_DEPTH_EXPERIMENT, stateSchema=11, headphoneFat=PENDING, publicRelease=false, dearVrParityProven=false; separate field guide SONKU_P94_FRONT_REAR_FAT.md included.

Operator A/B: compare against approved P9.1/P9.2 checkpoint 763706a1 or GUI-only P9.3 3648cee, at level-matched headphone playback, same routing/device and same .smix scene. Static 0, ±90,180 plus 6–12s orbit and yaw. Evaluate perceived front/rear distance and front/back confusion with vocal, bass and transients; reject if merely quieter/reverberant or loses punch. No production merge and no dearVR equivalence claim.

Copy link
Copy Markdown
Owner Author

MILESTONE A: archived / CLOSED UNMERGED. Historical private P9.4 front/rear A/B research, superseded by listening-approved SonKu P9.9 on main #733; do not replace the correct elliptical stage.

Original source branch research/sonku-p94-front-back-depth-20261116; original head 766b8d722ac18354e5c8e6eaa4e81f40ba42bcc8; original target feat/sonku-p93-circular-stage-movable-power-presets-20261116. Recorded with all 47 refs in docs/governance/MILESTONE_A_PR_BRANCH_RECONCILIATION_20261117.md (branch chore/milestone-a-pr-history-20261117). This is not a merge and not a claim that every prototype change shipped. Keep source/checkpoints exactly intact; no branch deletion, force-push, DSP tuning or release change. Reopen only for a specifically approved new scope.

@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