Repository navigation
perf(harmony): P0–P2 realtime pitch, phase-safe analysis and rate-stable tone - #487
30 commits merged into
Conversation
…amps and ASIO block parity
|
/arsonmix-ci fat ed11632 |
Deploying arsonmix with
|
| Latest commit: |
f02f35e
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://05c3bc32.arsonmix.pages.dev |
| Branch Preview URL: | https://feature-harmony-p0-bounded-r.arsonmix.pages.dev |
P0 exact-SHA qualification — 2026-11-09
|
… with stereo PCM regressions
|
/arsonmix-ci fat db33b3c |
Verified P1 checkpoint (exact SHA
|
…vals with tonal parity regressions
|
/arsonmix-ci fat b6e7968 |
2026-11-09 P0/P1/P2 exact-SHA qualification checkpoint
|
… and hostile-PCM safety
|
/arsonmix-ci fat b83597b |
2026-11-09 P0 realtime slice + hostile PCM milestone — exact HEAD verifiedCommit: Implemented: Reduced periodic autocorrelation work from 32 to <=8 candidate lags in a single sample, with analysis cadence sped from 8 to 32 slices per detection hop to avoid dropping frozen PCM windows. Added audio-owner completed/abandoned search counters and 32/44.1/48/96/192 kHz × 64/127/256 block regressions. Independently enumerated all integer sample rates 32,000..192,000 Hz: at least 62 samples of search-completion headroom (mathematical schedule, not measured ASIO timing). Stereo analysis-only fold now bounds its actual returned sample; hostile finite FLT_MAX, NaN and +/-infinity are safely contained before the voice's histories. No GUI, graph authority, VST routing, state packet or 48k DSP PDC change. Qualification: Fast native checks run 42581079521 PASS; Windows MSVC, native CTest, FAT integrity and Inno Setup run 42581191745 ALL SUCCESS. Downloaded artifact ZIP CRC PASS for both. Embedded installer and FAT provenance GitHub installer artifact: ArSonMix-Windows-Installer-Preview-42581191745 (ID 13658290036, 5,556,724 bytes ZIP; contains |
…ss graph/device changes
… and lost/recovered FX without reopen
…atic open-gate assertion
…fter automatic rebind
|
/arsonmix-ci ui 770f390 |
|
/arsonmix-ci fat 770f390 |
2026-11-09 field screenshot — Harmony GUI cannot control live DSP (P0 bridge recovery)Reproduced root cause from screenshot and source: UI footer said Fix Tests: New real Chromium Separate audio-quality defect: robotic voice is not resolved by this bridge fix; granular harmony shifter lacks formant preservation and 48k PDC is still 1536 samples (32ms). It needs independent resynthesis/formant and ASIO physical acoustic verification; do not equate GUI test success with global top-quality singer FX. |
|
/arsonmix-ci ui 4fd5d63 |
…backing and Space listening QA
2026-11-09 P0 Harmony recovery — native resynthesis and inspectable vocal A/B milestoneExact candidate HEAD: Actual code delivered:
Verified checks: exact-head Linux C++ native tests (all three) PASS 42627679146; fast approved UI and baseline native contracts PASS on same SHA. Windows full MSVC/CTest + verified FAT and Inno installer is running as one SHA-pinned canonical workflow 42628304651, triggered without changing this candidate via one-shot CI branch 42628287177. Do not declare Windows result before completion. Still P0 unqualified: source positive peaks are approximate waveform epochs, not proven glottal closure detection; retained grains/soft fade are engineering fallback, not certified formant preservation or natural vocal quality. No genuine adult singer male/female vowel/fricative/slides A/B performed here, no F1/F2/F3 report, no 30-min 48/96k ASIO xrun/roundtrip/CPU proof. Installer if successful is an experimental FAT only. Do not merge or claim global-class human voice until acoustic and physical gates; do not overwrite concurrent Player/Compressor/Throw work on main. |
2026-11-09 exact-HEAD Windows qualification COMPLETE — native CTest + Inno FATSource SHA
Still DRAFT and acoustically unqualified. This is a clean native engineering / source and packaging milestone, not evidence of a global-class human vocal sound: real adult singer A/B, vowels F1/F2/F3, breath/fricative intelligibility, pitch-slide expression, and physical ASIO 48/96k XRUN/roundtrip/callback p95/p99 remain outstanding. No merge, main modification, production release or hardware certification implied. |
…s, fixed ear timing and mono-safe depth
2026-11-09 D2 — Virtual backing-stage spatial enhancement (FIELD listening follow-up)Field input: previous pitch-synchronous Harmony is noticeably less robotic and already sounds reasonably natural. Remaining request is perceptibly greater spatial distance/3D separation between an intelligible forward lead and three musical backing singers, not more artificial pitch processing. Exact candidate HEAD: Source implementation
Qualifications
|
…dows CTest stack overflow
D2 Windows-specific CTest regression found and corrected (exact head 94a4039)The first full Windows run #42633886160 compiled successfully but had 69/70 CTest PASS, with a SegFault in Root cause from source + comparison with the previously green fixture: Repair: commit After fix: exact-head native C++ fast #42636728287 PASS; fast UI/contract checks PASS. One source-pinned complete Windows MSVC/70 CTest + Inno FAT requalification is now running: #42636832559. Do not use the older failed build or claim a new installer until its full result is confirmed. |
2026-11-09 Harmony P1 — singer-reviewed musical stage-depth refinement (exact SHA
|
Verified D2 spatial-stage Windows baseline (exact SHA 94a4039)Canonical Windows run #42636832559: SUCCESS. The SHA-guarded source Qualified experimental artifacts at this exact baseline SHA (14-day GitHub retention): Important live-head distinction: Harmony branch advanced afterward to |
…nup with offline A/B
D3 — Musical Room Focus, exact source
|
…sor main Two-parent integration preserves original Harmony history and current main Player/Compressor. CMake merges tests and dependencies additively. Retain PR Draft pending singer listening and physical ASIO.
Harmony D3 reconciliation checkpoint — 649409aPerformed an explicit two-parent integration on PR #487 branch only (not on main; no forced rewrite).
|
…ns clear of knob halo
Harmony D4 user playback fix — sweet-spot WET presence + dial label spacingOwner screenshot showed 4 voices, WET 100%, Space 60%, backing still too thin despite human-sounding timbre; knob captions COLOR/SPACE/WET/GAIN touch the arc. Exact branch HEAD:
|
Harmony D4 final UI+DSP reconciliation checkpoint — exact SHA
|
CHECKPOINT / HANDOFF — Owner field defects after Harmony D4 listening (2026-11-10 WIB)Owner audio acceptance (preserve): Vocal Harmony pitch/voice is now considerably more human, with D4 backing WET presence and stereo depth judged "lumayan bagus". Do not undo this acoustic baseline. Two new P0 defects reported by owner in actual application:
Live GitHub refs verified for handoff: Harmony PR #487 remains DRAFT at exact approved D4 checkpoint Source-backed P0 focus (not yet root-cause proof):
Required evidence / acceptance (developer thread must IMPLEMENT, not just propose): Native exact-input OFF/ON/OFF repeat at arbitrary sample offsets and 64/128/256 chunks, three/four/five voices, 48/96k minimum, 192k stress; same pre- and post-FX dry PDC aligned, no spike, drop-to-zero, tail burst, stale echo; rapid toggles reversals and rack OFF preserve individual FX state. Music source stable key, transitions C major/A minor, ambiguity, silence, disappearance, retarget and return, while vocal keeps singing: no click or drop, no stale Chapter/ChannelId, no accidental audible music WIRE. Automated PCM derivative/transient and spectral click evidence with comparison to prior baseline; finite/non-clipping, deterministic block-size independence, fixed latency. Measure DSP p50/p95/p99 and real ASIO4ALL driver XRUN counters separately, excluding false acoustic vs xrun classifications. No malloc, IO, locks, full-spectrum scans or logs in RT callback; one native PCM graph, no new audio thread. CI selective, exact SHA Windows CTest and FAT installer only after stable targeted tests; field user listening and physical ASIO required before promoting production. Action: Another thread should audit live refs, reproduce two mechanisms, fix architecture (not compensatory UI tricks), add deterministic regression fixtures and report PR/commit/tests/artifacts. Keep #487 draft and protect Player/Compressor/ASIO cross-thread work until evidence. |
P0 milestone after Harmony #467 and optimization #485
043748a552e66fd72e217cbf66aa35e6cc3e0b21(Compressor Compressor P0: realtime scalar optimization + PCM parity and multi-instance diagnostics #484 retained, no shared-file conflict).Acceptance
Exact HEAD: fast native+UI, complete Windows 67+ CTest, one optional experimental FAT package; no duplicate CI runs. Keep Draft and do not claim production-ready until field ASIO and recorded-vocal quality gate.
Supersedes #485 after qualification; preserve its tested CPU changes.
P1 phase-resilient stereo analysis (2026-11-09)
db33b3c66e6a0181a044356e1ccb06ef36f34318: added fixed-stateStereoPhaseSafeFoldto both vocal engine and read-only Music Track guide. Strongly anti-correlated stereo no longer erases pitch/chroma input; phase-coherent stereo retains original fold, 20ms evidence and 15ms smooth recovery, no new graph/router/PCM authority.ArSonMix-Windows-Installer-Preview-42570084359andArSonMix-Windows-FAT-Preview-42570084359. Both ZIP CRC verified after download; their embeddedsourceCommitequals the exact head SHA.2026-11-09 P0/P1/P2 deterministic DSP follow-up —
b6e796891e6602d4d7d9b15a9e1c5bf8427c2914RateStableColorlookup precomputed off RT. 48k settings keep legacy one-pole sound; 32/44.1/96/192k filters hold equivalent physical-time attenuation for the same Color value. Test checks six-sample/125 μs decay at 48k (not a too-long 10ms window that masks all differences).pow()computations per pitch analysis with 49 precomputed semitone intervals, one fractional basepow(), and bounded multiplies; saved a separate Color coefficient calculation per active voice. 42,600 semitone/detune ratio cases compared to the former formula within 1e-6. Pure C++ header, no new network/render owner or realtime allocations.2026-11-09 D0/D1 — Natural Backing Vocal Depth (exact-source working candidate)
8c75946d0081a81f52d8a4b75b37e30078f95686: remove Wet-dependent lead attenuation (~3dB at Wet 45%); isolate backing-only spectral shading, lower backing direct-to-reflected ratio via existing Space, preserve per-singer early taps, introduce one fixed-size shared 4-branch Hadamard FDN with prepared 0.55s nominal decay. No user-facing extra mode, schema change or new audio callback owner; plugin PDC unchanged.