Skip to content

[HEADPHONE FAT REJECTED] SonKu P8.3 native speaker research — music fidelity regression - #701

Closed
masarray wants to merge 26 commits into
research/sonku-p83-speaker-externalization-20261114from
feat/sonku-p83-physical-speaker-mono-stereo-20261114
Closed

masarray wants to merge 26 commits into
research/sonku-p83-speaker-externalization-20261114from
feat/sonku-p83-physical-speaker-mono-stereo-20261114

Conversation

@masarray

@masarray masarray commented Nov 14, 2026 •

Copy link
Copy Markdown
Owner

⛔ PHYSICAL HEADPHONE FAT REJECTED — 2026-11-15

Owner's actual native Windows mixer test: P8.3 sounded very quiet, hollow/boxed-in, constrained and substantially worse than P8.2. This verdict overrides all previous software CI passes. The Windows P8.3 portable candidate is NOT APPROVED for monitoring. Do not merge, publish, redistribute as qualified, or describe it as approaching dearVR. The approved baseline remains P8.2.

Proven source defects: positionedMode disables transparentBass, musicSafe, and sideRescue, removing the exact music-critical components that gave P8.2 its acceptable bass, MID integrity and stereo body. The P8.3 scene also does not receive P8.2's room presence law. Two independent HRTFs and the nonphysical LF480 2x2 transfer were accepted using four sparse frequency probes and lenient ±6dB criteria; directional ITD/ILD PASS says nothing about a music mix retaining its energy. Previous native PCM at 48k already measured stereo anti-phase total energy: P8.2 8.9678, Physical 1.5209 (−7.71dB RMS), LF480 5.4893 (−2.13dB RMS), plus LF480 −5.801dB at 461Hz and −3.887dB at 1703Hz in separate test cases.

Containment committed: add source-level stricter ±2.5dB multiband and ±3dB broadband program fixture vetoes; don't conceal loss through auto-gain. Source CI must FAIL the existing poor candidate. P8.3 full Windows workflow is manual-only with the source veto before expensive JUCE build. No new portable until real fix passes all source gates, then owner headphone A/B. Golden P8.2 and main GUI are preserved. See docs/SONKU_P83_NATIVE_MIXER_FIELD_FAT_20261115.md.

Next development must be architectural, not a makeup-gain/reverb patch: design two-ear stereo transfer with measured localization cues and constrained music-spectrum/phase/coherence preservation, make MonoObject robust, calibrate direct-to-room balance and test full-band musical content plus blind perceptual direction. Generic KEMAR and head-fixed physics still limit externalization. Human listening remains final gate.


SonKu P8.3 — native world-fixed speaker research, LF music compatibility and real PCM A/B (DRAFT)

Baseline and integration safety

This PR is stacked on #697 -> #695 (P8.2), NOT on the newer production main GUI. It preserves P8.2 as the golden, operator-tested Music Bed; the prior P8.2 tests showed a center KEMAR direct MID with azimuth-invariant unpositioned SIDE headphone bed. Physical P8.3 speaker mode is a separate experimental transfer. No changes to native PCM graph authority, /preview/ GUI, public .smix, or shipped DSP defaults. No merge, release or dearVR claim without owner headphone FAT.

Implemented native algorithm

  • New MonoObject and independent SpeakerMonitor L/R 2x2 measured KEMAR sources, default positions ±30° FRONT; direct source and six-surface image-source reflection coordinates are world-fixed under deterministic head yaw.
  • Exact native direct/horizontal/floor/ceiling/late-only PCM stems; evidence for path delay, ILD, ear-correlation, room DRR and combined-stem reconstruction.
  • Two EXPLICIT, research-only stereo speaker variants: (A) physically measured 2x2 HRTF, which can legitimately cancel anti-phase LF at a symmetric listening point; (B) Music-LF480 with complementary 480-Hz low-band direct 2x2 head-relative ear conditioning. The latter is not a physically measured low-band BRIR and trades some localization physics for music fidelity. No generic unpositioned full-band headphone SIDE bed, auto-coherence switching, gain rescue, or room-reverb inflation. MonoObject does not use the LF compromise.
  • LF sweep tested 150/220/350/480/650 Hz, selecting 480 Hz as the lowest tested that meets four test-band criteria at 44.1/48k. Original physical speaker mode remains the audit reference.

Measured source evidence (latest tested DSP source commit 8bee0e899b3f890bcbea46e81d49bd30493c1798)

Fast native C++ CI PASS: P8.2 Music Bed is bit-identical across independent P8.2 and P8.3-compiled builds, 44.1/48 kHz ROOM ON/OFF and mono/antiphase. Acoustic geometry, yaw/wrap, eight cardinal/diagonal directions, isolatable room stems and 64/128 block parity PASS. Opt-in Music-LF480 anti-phase vs P8.2 at 48k: 83 Hz -1.693 dB, 461 Hz -5.801 dB, 1703 Hz -0.173 dB, 5387 Hz +0.917 dB; mono: +1.053 / +1.096 / -3.887 / -2.781 dB respectively. Four-bin ±6dB guard passes on both 44.1 and 48k for mono and anti-phase (not sufficient to prove total tonal quality).
Fast CI run https://github.com/masarray/arsonmix/actions/runs/43631205592 ; diagnostic WAV artifact ID 14127519462. The previous pure physical P8.3 anti-phase bass failure (-25.112 dB at 83Hz) remains an expected physical property, not secretly renamed as a physical PASS.

Usable owner A/B deliverable

tools/sonku_p83_offline_ab.cpp runs the ORIGINAL SonKuSpatialCore on user-supplied stereo WAV (PCM16/24/32, float32 and common WAVEFORMATEXTENSIBLE; 44.1/48k) and produces three exact-input stereo IEEE WAV files: P8.2 Music Bed / P8.3 physical / P8.3 LF480 music-compromise. ROOM OFF is default, optional ROOM ON. Native CI compiles it and smoke-tests standard PCM16 + DAW extensible float32 for room-on/off, verifies output WAV layout, uploads three rendered A/B examples and RMS/peak CSV. No input audio is uploaded to external services by the utility. It is offline ONLY, not a new ASIO/native JUCE output path, and outputs are intentionally not loudness-matched.

CPU and CI efficiency

Hosted Linux 48k/64/4-insert latest measured: P8.2 p50 642.11us p99 702.03us; P8.3 LF480 p50 665.41us p99 808.68us; both had 0 hosted deadline overruns in this ONE run. Not ASIO XRUN or real audio hardware proof. Bench also covers 44.1/48k 64/128/256 and 1/2/4 instances.
Three inherited P7.4/P7.5/P8.0 historical Windows FAT workflows were incorrectly triggering full JUCE builds on every P8.3 source push; changed them to manual-only workflow_dispatch, retaining the project's selective milestone FAT policy. The latest commit 659894137e6ee09299ea3ea5916285d517e3964d changes only those CI controls and the engineering documentation after the verified DSP source commit.

HARD acceptance gates before portable Windows P8.3 JUCE or public release

  • Confirm real headphone externalization, 4/8-direction blind positioning, front/rear/overhead errors, and level-matched P8.2/dearVR A/B from owner listening.
  • Further inspect full-band music spectra (not just 4 bins), harmonics, transient/kick/bass, time/phase response, DRR and low-band position compromise, and optimize core against real measured headphone profile.
  • Integrate research render-mode and head rotation into the actual JUCE adapter with versioned .smix / SPSC audio-applied ACK, rollback/state restore, device restart, and minimal accepted GUI, while keeping the one PCM graph. This has NOT been done.
  • One selective exact-head native Windows FAT/portable candidate only AFTER native mode integration, not one of the historical P7 Windows ZIPs. Physical ASIO FAT remains PENDING.

Track issue #683. See docs/SONKU_P83_NATIVE_PHYSICAL_SPEAKER_CHECKPOINT_20261114.md for reproducible exact-source evidence, WAV input guidance and limitations. Keep DRAFT.

…ed preview (#690)

Preserve the approved operator UI. Add static brushed-steel PNG with slimmer cap, 30-segment LED ring and fixed green pointer, true-black LCD, contextual vertical GAIN label, and focused three-viewport Chromium QA.
… detent (#694)

Preserve current approved /preview/ mixer. Layer LED behind static steel cap shadow; crop baked-in rim; use a true inset black LCD with lowered digits; share zero detent across browser and native controls, ensuring exact native ACK/readback and wheel commits; validate with scoped Chromium 3-viewport suite.
…696)

Prevent transparent-slider click-to-seek, add 6px relative drag intent and 0 dB capture/release hysteresis, preserve native ACK/readback, and apply requested fixed shadow and inner focus halo. Verified 3 Chromium resolutions on approved /preview/.
#698)

Preserve approved click-safe 0 dB detent and raised metal hardware. Scale dial to 58px and LCD to 84x25px with responsive CSS tokens, quieter 4px LED ring, and five-viewport Chromium visual/interaction checks. No DSP/PCM/native bridge or session changes.
…dance (#700)

Reference exact GitHub safety checkpoint #699 from existing AGENTS.md, warn about stale integrated GUI PRs, and keep future reconciliation at latest-main file hunks. Documentation-only, no runtime/CI/Pages/PCM edits.
@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: e40e723
Status: ✅  Deploy successful!
Preview URL: https://b9967523.arsonmix.pages.dev
Branch Preview URL: https://research-sonku-p84-musical-c.arsonmix.pages.dev

View logs

…le (#702)

Restore speculative LCD/needle to accepted ChannelId value after unchanged-value rejection/timeout, avoid redundant DOM painting and full strip rebuild on gain-only error, cancel gain drag on lost capture, blur or mode redraw. 5-view Chromium and fast source/native contracts PASS. No PCM DSP or visual CSS changes.

Copy link
Copy Markdown
Owner Author

P8.3 source-level checkpoint — exact head 5252da9e78fccd18042b8c4c4cfe1c01d46b3c3f (2026-11-14)

Implementation: research-only MonoObject and true 2x2 L/R SpeakerMonitor FIR at ±30°, world/head yaw separation, world-fixed six-wall reflection directions, native isolated room stems, no new device callback, no public GUI/.smix change. Branch based on research #697 / golden P8.2 #695, not latest main.

Verified fast-native evidence: Native C++ source/geometry/ear signs/room-stem reconstruction PASS at 44.1 and 48 kHz; first-order reflection combined stem reconstruction max error 2.91e-11 at 48 kHz. Two separate source compilations, P8.2 with/without P8.3 research macro, have bit-identical 44.1/48-kHz original music PCM for ROOM ON/OFF and correlated/antiphase inputs. Existing P8.2 known-defect diagnostic remains PASS as a negative-control reproduction.

Audio quality veto (intentional FAIL): At 48 kHz, P8.3 vs P8.2 anti-phase 83 Hz is -25.112 dB, at 1703 Hz -1.374 dB. Normalized LF-versus-mid gap -23.738 dB, exceeding the -6 dB music-integrity gate. Similar at 44.1 kHz (-26.095 dB vs -2.113 dB). Source passing directional ITD/ILD is not a perceptual music pass. Unmodified P8.2 remains the safe listening baseline. Do not conceal the failure with gain, extra reverb, forced elevation or a disguised unpositioned headphone-side bed.

Hosted Linux CPU preliminary evidence (not ASIO): 48k / 64 frames / 4 inserts P8.2 p50 746us p99 1092us, P8.3 p50 778us p99 1553us, with 3/64 hosted blocks over 1333us; non-deterministic shared-runner scheduling prevents real XRUN claims. Other 44.1/48k 64/128/256, 1/2/4 combinations were logged. Fast source gate compiles and executes, P8.2 parity and hosted CPU steps PASS, P8.3 fidelity gate intentionally FAILS. Exact-run logs: https://github.com/masarray/arsonmix/actions/runs/43616937803 ; offline source PCM artifact ID 14122238679 (not approved music or physical FAT).

Release veto: No Windows portable P8.3 research candidate dispatched, no native .smix/ACK/UI feature integration, no merge or public release, no dearVR parity claim. Next work must solve the LF speaker-transfer/music-fidelity conflict with objective spectral and headphone evidence, then run selective full native Windows FAT on that qualified head. Keep this PR DRAFT.

@masarray masarray changed the title SonKu P8.3: native physical mono object and world-fixed stereo speaker PCM gates SonKu P8.3: world-fixed native speakers, music-compatible LF A/B and verified PCM diagnostics Nov 14, 2026

Copy link
Copy Markdown
Owner Author

SonKu P8.3 meaningful milestone checkpoint — 2026-11-15 (Jakarta)

Latest PR head 659894137e6ee09299ea3ea5916285d517e3964d is CI/workflow+docs only after exact-source DSP/AB 8bee0e899b3f890bcbea46e81d49bd30493c1798. Still DRAFT on #697 research baseline, no merge or public release.

Source evidence PASS: Native original C++ SonKuSpatialCore, 44.1/48k, PCM/multi-band/phase/directional room stems, combined decomposition, source+speaker yaw, Music Bed bytewise preservation, bounded block sizes. LF compatibility fixed the explicit anti-phase 83Hz -25.112dB defect to -1.693dB (48k, vs P8.2), and 461Hz to -5.801dB (vs -11.520dB at 150Hz crossover), with independent mono and anti-phase four-band ±6dB gates on both sample rates. SpeakerPhysical retains true measured pair/expected anti-phase cancellation; LF480 explicitly sacrifices physical low-mid accuracy. No headphone externalization proof.

Usable owner audition: tools/sonku_p83_offline_ab.cpp renders own licensed stereo WAV with the actual SonKu core to P82 Music Bed / P83 SpeakerPhysical / P83 SpeakerMusicLF480, ROOM OFF or ON. 16/24/32-bit PCM, float32 and DAW WAVEFORMATEXTENSIBLE smoke-compilation PASS, real WAV header verification PASS. Outputs raw relative levels and RMS/peak CSV; level-match before human comparison. Latest source CI + downloadable original synthetic WAV artifact: https://github.com/masarray/arsonmix/actions/runs/43631205592 (artifact ID 14127519462, generated from PR head 8bee0e8). No native Windows exe was requested for P8.3 and no full JUCE live mode/ACK/.smix has been implemented.

Hosted CPU limited comparison: 48k 64 frames 4 inserts, P8.2 p50/p99 642/702 us, P8.3 LF480 665/809 us in one shared Linux runner measurement, 0 simulated blocks over 1.333ms there; no ASIO/XRUN claim.

CI budget fix: three historical P7.4/P7.5/P8.0 pull_request workflows were still launching duplicated full Windows builds for normal P8.3 SonKu header changes. Converted these historical FAT workflows to explicitly dispatched only. No need to run older P7 Windows FAT ZIPs for P8.3. Use one deliberate exact-SHA Windows native FAT when the host's P8.3 research mode + SPSC/ACK and versioned session are implemented, then owner headphone A/B. No public merge until owner approves music and realism.

Open sound/engineering risks: measured true-speaker stereo anti-phase LF null is physically expected; 480Hz music guard introduces nonphysical ILD in the bass/lower mids. Four-bin pass is insufficient for dense mixes, pinna cues/front-back, room plausibility, or externalization; generic KEMAR and no live head tracking are limitations. Do NOT claim dearVR parity, binaural externalization or hardware ASIO correctness from CI.

Copy link
Copy Markdown
Owner Author

P8.3 actual native mixer Windows FAT — first compile & fail-closed regression (15 Nov 2026)

  • First exact-source candidate 79f307fa227f2fda857a1364c6271f8087a60d63 successfully compiled the real ARMixConsole Windows .exe and JUCE P8.3 test target, not an offline mock.
  • Scoped CTest: 7/8 PASS, 1/8 FAIL. NativeSonKuP83JuceV4AudioAckLegacyRestoreAndHeadYaw discovered a genuine malformed FX-state acceptance: juce::MemoryInputStream::readByte() can yield signed -1 for 0xff; testing only >1 fails open for both new v4 low-frequency flag and inherited v1-v3 motion/direction bool. Evidence: https://github.com/masarray/arsonmix/actions/runs/43634359115. Packaging correctly did not run, no test candidate was distributed.
  • Fix commit ed74277c750bcdc6393ccd38fb49257e12ee00ea requires boolean bytes exactly 0 or 1, adds a real logical/unpublished mode edit completing without a nonexistent audio callback, and guards session quiescence with an atomic pending-mode flag through a single native 12ms fade transaction. The JUCE test now explicitly seals its real-time graph, checks ACK is delayed until PCM application, verifies logical edits, and corrupt-state denial.
  • Corrected exact-SHA P8.3 Windows candidate build triggered at https://github.com/masarray/arsonmix/actions/runs/43635510105. Its fast PCM+GUI preflight PASS; Windows app+8 focused tests/packaging still subject to verification. Do not merge/release, and do not claim ASIO/headphone externalization until owner's FAT.

Maintains last-approved P8.2 Music Bed bytewise source invariant, updated latest-main GAIN GUI, public P8.3 compile flag OFF, and draft PR #701 status.

Copy link
Copy Markdown
Owner Author

P8.3 PHYSICAL HEADPHONE FAT CANDIDATE READY — full native Windows JUCE verified (15 Nov 2026 WIB)

Single exact-source artifact: ArSonMix-SonKu-P83-Native-Mixer-Windows-x64-43635510105 (GitHub Actions artifact ID 14129764387, archive 15,102,256 bytes, not expired). Linked workflow #43635510105 conclusion SUCCESS, exact branch HEAD ed74277c750bcdc6393ccd38fb49257e12ee00ea.

Windows full build: actual ARMixConsole application AR MIX Studio.exe (15,046,656 bytes) compiled successfully with P8.3 source flag ON, v4 native JUCE adapter and latest main GAIN GUI. Scoped 8/8 Windows CTests PASS, including live-PCM mode edits, actual v4 SPSC audio-applied ACK/legacy v1-v3 restore, P8.2 room / P8.1 tonal, P7.2 native. Packaging PASS, BUILD_PROVENANCE.json source SHA identity PASS, GUI/PE resources exact and static WebView2 startup checks PASS, MIT KEMAR notice and SONKU_P83_FIELD_FAT.md included and checked. GitHub artifact archive SHA256: 7c8116f7324cd461ff8d17c63f303d485e9412b6cd3c6c676b203dd4edf5e00a. This is a research portable, not public release.

Quick real mixer FAT: extract all files separately from previously approved P8.2; run native mixer EXE; set real headphone audio output (48 kHz/128 frames as initial test) and no Master Mono or unwanted duplicate dry MAIN path; load stereo music through actual channel → 8D-SonKu; open SonKu native editor and confirm real schema-v4 RENDER MODE selector after identity-bound native snapshot; compare Music Bed P8.2 / Speaker Physical / Speaker Music LF480 (not physically measured LF BRIR), with same source, speaker ±30° front/0° elevation/1.6m/Room OFF first. Confirm audio APPLIED/readback before trusting mode. Test head yaw and direct/room plus blind front/back/overhead and music/phase; only owner hearing can approve. Full guide included SONKU_P83_FIELD_FAT.md and source docs/SONKU_P83_NATIVE_MIXER_FIELD_FAT_20261115.md.

Vetoes remain: no physical ASIO XRUN/fidelity result yet, no actual human externalization/dearVR equivalence proof, HRTF generic & LF480 nonphysical. PR #701 remains DRAFT, main and public release unchanged, golden P8.2 retained for A/B.

@masarray masarray changed the title SonKu P8.3: world-fixed native speakers, music-compatible LF A/B and verified PCM diagnostics [HEADPHONE FAT REJECTED] SonKu P8.3 native speaker research — music fidelity regression Nov 14, 2026

Copy link
Copy Markdown
Owner Author

HEADPHONE FAT FAIL — 15 Nov 2026: User compared the actual full native mixer P8.3 portable and reported it as much quieter, hollow, constrained and plainly worse than P8.2. P8.2 remains golden. Prior green CTest is not acoustic approval. Source-level cause: positionedMode bypasses P8.2 transparent bass, music-safe MID/side and SIDE stereo restoration; numerical 48k anti-phase output energy confirms severe loss (P8.2 8.9678, physical P8.3 1.5209, LF480 5.4893). Updated draft PR description; new music fidelity source hard veto at commit e40e72367c0e7a7a67cc463efe7c32564daa8aaa and full Windows P8.3 workflow manual only with source preflight. Existing artifact from run 43635510105 is rejected forensic reference, NOT an approved test/monitoring release. Do not merge/release until original musical fidelity and externalization pass independent headphone FAT.

Copy link
Copy Markdown
Owner Author

Milestone A — archival closure (NOT MERGED). This experimental SonKu P8.3 branch was explicitly rejected by the operator's physical listening test (see this PR's rejection evidence). The selected, listening-approved production SonKu P9.9 was later integrated through #733; it is not authorized to reapply rejected P8.3 processing. Original experimental head e40e72367c0e7a7a67cc463efe7c32564daa8aaa, source branch feat/sonku-p83-physical-speaker-mono-stereo-20261114. The exact SHA and boundary are recorded in docs/governance/MILESTONE_A_PR_BRANCH_RECONCILIATION_20261117.md on branch chore/milestone-a-pr-history-20261117. No merge, code change, force-push, checkpoint move, or branch deletion. Reopen only for a newly scoped, separately approved task.

@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