Skip to content

SonKu P8.3 research: diagnose in-head stereo SIDE and design true external virtual speakers - #697

Closed
masarray wants to merge 3 commits into
feat/sonku-p82-sixsurface-multiband-room-20261114from
research/sonku-p83-speaker-externalization-20261114
Closed

masarray wants to merge 3 commits into
feat/sonku-p82-sixsurface-multiband-room-20261114from
research/sonku-p83-speaker-externalization-20261114

Conversation

@masarray

Copy link
Copy Markdown
Owner

Owner's verified listening

P8.2 tonality and room sound good, but still head-locked and at times above the head, not a convincing FRONT/RIGHT/REAR/LEFT speaker. This is NOT a request for stronger reverb.

Source findings

  • P8.2 DIRECT uses one KEMAR center HRTF for musical MID plus a retained stereo SIDE bed independent of source azimuth. This protects phase but is not the binaural transfer of speakers positioned around the listener.
  • ROOM models first-order six surface reflected L/R sources, with different geometry from the centered direct source. This incoherence can degrade externalized object impression.
  • 0° direct/elevation means ear height geometrically; ceiling/floor ER may contribute perceptually to reported 'above head' but this mechanism is hypothesis until isolated listening/stem diagnostics.
  • One generic KEMAR, no head pose/world-lock in native source; literature confirms front-back/elevation and head motion issues are common. Raw 48k KEMAR +90 right-ear energy > left, so don't swap channel sign to attempt a fix.

This PR intentionally changes NO shipped DSP

  • P8.3 engineering RFC, grounded in Best et al. externalization review, Brimijoin head movement, 2026 generic HRIR study, frontal BRIR reflection experiments, Dear Reality Clarity, Steam Audio architecture and Waves Nx speaker/head-tracking workflow.
  • Native tests/sonku_p83_externalization_gap_diagnostic.cpp compares FRONT/RIGHT/REAR/LEFT on the exact P8.2 engine, captures horizontal ear ILD, and intentionally confirms direction-invariant anti-phase SIDE with room OFF, 44.1/48 kHz. This is a reproduction of a defect, not a sound-quality PASS.
  • Fast Linux PCM CI only; no unnecessarily slow native Windows installer rebuild before redesign exists.

Proposed P8.3 future stages

  1. P8.3A direct/ER/ceiling-only/late stems; per-band ITD/ILD, elevation/front-back and correlation to isolate actual flaw.
  2. P8.3B distinct Speaker Monitor mode using true 2×2 BRIR/HRTF speaker mapping (±30°, plus discrete monophonic positions around head), coherent frequency/phase safety and per-source reflections. Preserve original P8.2 as separate MUSIC mode.
  3. P8.3C optional head-tracked world-fixed scene and legal HRTF profile/headphone calibration after native speakers are verified.
  4. P8.3D real headphone localization confusion matrix, loudness-matched dearVR PRO 2, ASIO p95/p99/XRUN and owner signoff.

Draft / release veto: no public merge, no claims of dearVR parity; current P8.1/P8.2 physical binaries remain unchanged. Engineering issue #683 remains open.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 14, 2026 •

Copy link
Copy Markdown

Deploying arsonmix with  Cloudflare Pages  Cloudflare Pages

Latest commit: df55c0e
Status: ✅  Deploy successful!
Preview URL: https://42361b48.arsonmix.pages.dev
Branch Preview URL: https://feat-sonku-p83-physical-spea.arsonmix.pages.dev

View logs

Copy link
Copy Markdown
Owner Author

P8.3 exact-engine externalization diagnosis: direction-invariant stereo SIDE CONFIRMED

Owner's physical verdict: P8.2 music/tonality sounds good, but sources are internalized or appear above head; FRONT/RIGHT/REAR/LEFT do not resemble real external speakers. This is NOT a request to increase reverb and does not withdraw P8.1/P8.2 tonal approval.

New dedicated Draft research PR #697: #697, branch research/sonku-p83-speaker-externalization-20261114, exact head df55c0e958b858906f7329b61fa2df7b11f5a2d6.

New native C++ source diagnostic CI #43610825163: https://github.com/masarray/arsonmix/actions/runs/43610825163 — SUCCESS with all full P8.2 compile flags. Importantly, SUCCESS means reproducibly demonstrated a defect, NOT speaker externalization passing.

Measured using actual P8.2 SonKuSpatialCore input with hard antiphase stereo (SIDE-only) and ROOM OFF:

  • At 44.1kHz, maximum sample delta across FRONT/RIGHT/REAR/LEFT = 0.0000000000; side energy nonzero. 48kHz = 0.0000000000 too. The content remains independent of azimuth and headlocked.
  • Mono source LEFT -90° vs RIGHT +90° yields +5.652 dB 44.1k / +5.849 dB 48k in the appropriate near ear; FRONT/REAR ILD = 0 dB. This rules out a global L/R sign swap as the obvious fix.
  • ROOM ON alters the response and varies with azimuth, but does not change the fact that the direct SIDE bed has no directional speaker transfer. P8.2 direct and room also represent differing stereo source geometries.

Leading architectural cause is now source-confirmed: P7.5→P8.2 direct uses centered MID KEMAR HRTF + highpass headphone SIDE, not a physical two-loudspeaker binaural transfer matrix. Additional contributors remain unverified hypotheses: generic KEMAR frontal/rear ambiguity, ceiling/floor ER and vertical interpretation, level/DRR, head-fixed field without head tracking, headphone EQ.

P8.3 strategy (documented at SONKU_P83_REAL_SPEAKER_EXTERNALIZATION_ARD_20261114.md):

  1. Isolate direct, horizontal ER, floor, ceiling, late separately and log per-band ITD/ILD/ER timing; do not assume overhead is caused by ceiling without an A/B.
  2. Add separate opt-in Speaker Monitor mode: coherent-safe physical 2-input×2-ear speaker HRTF/BRIR matrix with consistent direct and wall images; keep existing P8.2 Music Bed intact.
  3. Optional listener/world head pose (timestamped and SPSC, separate from source orbit) and legal selectable HRTF/headphone profiles; no sensor/OSC I/O on PCM callback, no dependence on unowned hardware.
  4. Physics/PCM invariants, explicit front/back/overhead blind headphone direction tests against dearVR with same settings, and ASIO 1/2/4 insert p95/p99 before qualifying sound.

Scope today: measurement test + research ARD only; NO P8.2 DSP modified, NO new Windows binary, no public release, no unproven dearVR parity claim. PR #697 Draft. Issue #683 remains open.

Copy link
Copy Markdown
Owner Author

MILESTONE A — research archival (closed, NOT merged). This P8.x SonKu research candidate is historical/experimental and is not the owner-approved P9.9 production path. The real approved main integration is #733, by selective forward-port, not by importing all earlier research commits. Original head df55c0e958b858906f7329b61fa2df7b11f5a2d6, source branch research/sonku-p83-speaker-externalization-20261114, original base feat/sonku-p82-sixsurface-multiband-room-20261114 are pinned in docs/governance/MILESTONE_A_PR_BRANCH_RECONCILIATION_20261117.md on chore/milestone-a-pr-history-20261117. No branch, tag, or checkpoint deletion; no binary/DSP/GUI/session change. Archive the review only, retain complete source evidence. Reopen if explicit new research is approved.

@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