Skip to content

DRAFT Harmonice P1 — isolated 32/24/16 ms latency experiment, PCM/PDC and aligned WAV A/B - #514

Closed
masarray wants to merge 18 commits into
fix/harmony-main-reconcile-dropout-bypass-20261110from
experiment/harmonice-p1-latency-audit-20261110
Closed

masarray wants to merge 18 commits into
fix/harmony-main-reconcile-dropout-bypass-20261110from
experiment/harmonice-p1-latency-audit-20261110

Conversation

@masarray

@masarray masarray commented Nov 10, 2026 •

Copy link
Copy Markdown
Owner

Scope: independent experimental P1 milestone, NOT a live latency switch

Stacked on #509 exact baseline 618cb67c476ffe41904175cc87890f075c9c3b21, which itself forward-ports only the 8ms bypass and Music guide stale-reacquisition fixes from current main. Do not merge this PR ahead of #509; do not merge older QA #499/#503/#506 lineages.

  • Shipping Engine::prepare(rate), the only path used by native ChannelHarmonyProcessor, retains its exact original 32ms nominal algorithmic latency. No default DSP quality or ASIO path changed.
  • An explicitly named offline-only preparation method explores 24ms and 16ms grain designs. The DSP physics, true latencySamples(), ring safety, time-domain retune and voice stage remain intact.
  • Native C++ test verifies all 32/24/16ms candidates at 44.1/48/96k with 64/127/256 blocks: sample-exact Wet0 impulse/PDC, bounded/finite active 5-voice output, same PCM across callback partitions, and anti-phase input. Synthetic PCM safety only — NOT acoustic approval.
  • tools/harmony_offline_ab.cpp --latency-study input.wav output_directory writes aligned 32ms reference, aligned 24/16ms mix/dry WAVs, and latency-study.json with actual frame counts/offsets; qualityAccepted=false and physical ASIO unverified. Use genuine original adult-singer WAV recordings for human evaluation.
  • Lightweight Harmony fast Linux job compiles native latency safety test; Windows complete CTest/ASIO listening still required before consideration.

Explicit exclusions

No change to routing, native mixer callback, JUCE/device/ASIO, SOURCE ONLY/WIRE, Player, Compressor, Vocal Throw (owner marked sound-quality blocker), GUI, session schema, or installer. No claims of perceptual superiority or callback speedup until same-runner measurements and singer listening gates.

Decision gate

Reject 16ms or 24ms if consonants/formants, low notes, stereo compatibility, CPU p99.9, or backing timing deteriorate. Do not activate candidate in production until proven. This draft is an inspectable test fixture and safer starting point, not a release candidate.

Measured preliminary exact-SHA qualification (2026-11-10)

Milestone D quality gate update — 2026-11-10

Current head ffe429ff992314ffbf1674469b783f487b19b00f, still stacked on #509. Linux exact-head Harmony fast #42756943860 PASS; new low-register causal PSOLA test + executable WAV report contract. UI smoke #42756943845 PASS. QA mirror #516 (never merge) was used to request one exact-SHA Windows full CTest via workflow #42757047991; inspect final result, do NOT infer pass from request.

Engineering quality result: 16ms has only 16ms causal lookback, but PSOLA at 65Hz requires 18.46ms (1.2 source period). Native DSP tests at 44.1/48/96k confirm zero PSOLA-ready samples at 65Hz for 16ms, while 24/32ms remain ready. Therefore 16ms REJECTED as full-range formant-safe candidate; 24ms only tentatively feasible, no singer A/B approval, production still 32ms.

Four-source generated SYNTHETIC corpus (65/220/440Hz and anti-phase), aligned WAVs and JSON metrics cover dry sample parity, wet-only RMS, mono/stereo, onset step and actual per-candidate PDC. Python validates actual float32 RIFF output, not simulated stubs. Run ARMixHarmonyOfflineAb --synthetic-corpus /folder or --latency-study real-original-adult-vocal.wav /folder. Full details docs/HARMONICE_P1_QUALITY_LATENCY_EVIDENCE_20261110.md.

No user-facing latency switch, ASIO alteration, GUI/source identity/session/WIRE change, or Vocal Throw merge. Physical singer, measured ASIO roundtrip/XRUN, true formant preservation, and 24ms perceptual quality NOT VERIFIED.

@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 10, 2026 •

Copy link
Copy Markdown

Deploying arsonmix with  Cloudflare Pages  Cloudflare Pages

Latest commit: ffe429f
Status: ✅  Deploy successful!
Preview URL: https://d55c8824.arsonmix.pages.dev
Branch Preview URL: https://qa-harmonice-p1-quality-exac.arsonmix.pages.dev

View logs

Copy link
Copy Markdown
Owner Author

Superseded by clean #520 which copies exactly the same nine P1 study files onto the branded, verified P0 #518 parent. #520 current exact SHA 34cec4e44434af8226f75db38cd088f79503796d: Linux Harmony fast DSP #42768671707, fast Windows contract #42768671630 and UI source contracts #42768671712 all PASS. The earlier Windows full source-study run 42757047991 passed 83/83 but was tied to original QA SHA and does not qualify the new head. The original branch's parent predates branded getName()=Harmonice; keeping this stacked draft would risk restoring obsolete display identity. Closing #514 UNMERGED, not promoting 16/24ms to production; use #520 for all subsequent listening/ASIO qualification. 16ms at 65Hz remains disqualified by causal PSOLA lookback. No physical singer/ASIO certification.

@masarray masarray closed this Nov 10, 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