research(mir): STFT vs CQT 구조 세그멘테이션 비열등성 검증

[seonghobae]

문제

#1223은 chroma_cqt → chroma_stft가 synthetic/mock 조건에서 빨라질 가능성만 보였고, 실제 디코딩 음원에서 구조 경계·기능 라벨·반복 그룹 품질이 보존되는지는 검증하지 않았습니다. BandScope의 structure analysis는 rehearsal buyer truth이므로 계산시간만으로 production representation을 바꾸지 않습니다.

검증 질문은 chroma_stft가 현재 chroma_cqt 대비 사전 정의된 구조 품질 비열등성 기준을 만족하면서 같은 실제 음원·runtime·execution host에서 재현 가능한 latency 개선을 제공하는가입니다.

Production acceptance

• synthetic waveform이 아니라 rights-cleared real decoded music corpus를 사용하고 license/provenance/audio SHA-256/annotation SHA-256을 고정합니다.
• 동일 admitted finite PCM과 exact runtime/host identity에서 baseline/candidate를 paired 측정합니다.
• duplicate audio_sha256은 독립 관측치처럼 세지 않습니다.
• schema v1에는 dropout/exclusion 허용치가 없으므로 실패 track은 PASS를 막습니다.
• claim boundary, noninferiority margin, latency threshold, aggregation, paired uncertainty, resample count/seed는 candidate 결과를 보기 전에 고정합니다.
• production switch는 complete registered experiment PASS 뒤 별도 PR에서 수행합니다.

구현 상태 — Draft #1228

Current exact head는 e99ccecc20a72d9531d77be5b854ae6090500aff, base는 protected develop@314ddeae7b775a4957594b599358c8255617eb2e입니다. PR은 Draft/Open/mergeable이고 protected base 대비 ahead 200 / behind 0입니다. Production structure extraction default는 여전히 chroma_feature="cqt"입니다.

Functional-label ACC는 pinned MIREX-2025 reproduction b9fa0b0...::calculate_accuracy의 200 ms grid를 사용합니다. Exact mir_eval==0.8.2 research overlay와 detection/deviation/pairwise arguments, runtime-lock digest가 scientific identity에 결합되어 있습니다.

aggregate_structure_noninferiority.py는 macro-track-v1 / paired-track-bootstrap-v1을 소유합니다. Registered track pair를 sampling unit으로 하며 equal-track macro aggregation, PCG64 paired bootstrap, NumPy linear 0.025/0.975 percentile interval을 고정합니다. Stored aggregate/CI는 complete track receipt로부터 canonical recomputation한 값과 exact-match해야 합니다.

isolated-single-shot-v1은 macOS/Windows에서 각 track·lane을 fresh subprocess에서 20회 single-shot 측정하고 same-process warm-up은 0회입니다. Trial 순서는 CQT→STFT / STFT→CQT로 교대합니다. Latency는 repository segmenter만 time.perf_counter_ns()로 측정하고 NumPy linear p50/p95를 구성합니다. Peak RSS는 worker lifetime maximum이며 macOS ru_maxrss, Windows PeakWorkingSetSize를 사용합니다.

Canonical execution host / experiment runner

Fresh review에서 registration.runtime.host_profile이 단순 non-empty text였고 실제 performance execution host와 비교되지 않는 gap을 확인했습니다. p95 latency ratio가 decision criterion인 만큼 host는 설명용 라벨이 아니라 scientific condition입니다.

Draft #1228은 이제 structure-performance-host-v1:<sha256>을 canonical execution prerequisite로 사용합니다. Payload에는 platform, OS release/version, architecture, hardware model, CPU model, logical CPU count만 포함하고 hostname/account/serial/MAC/IP/path는 제외합니다. macOS는 fixed sysctl selectors, Windows는 fixed machine-level registry values를 읽습니다. 실제 captured profile hash가 registration과 다르면 corpus admission 전에 fail closed합니다.

RED 7a90b922c0f64090f1ac2717e9afc0369503bd58 → GREEN 70b86022323728c6b3f85d5f14468423ebf15478으로 host binding과 canonical runner를 추가했고, RED f1a6971ff7f60af993c05455e40ccc6d57ab593e → GREEN a31bebd79d88b940c9f96517d02f7c0364fac485 / 7b07c00d9ee8c7ddbb24f3a65bd62aa1e22c05c4로 free-form host label을 canonical execution에서 거부하도록 고정했습니다. 7d3c28d55af1b1ba08afc9ed2a6df85274a7d6f7에 execution/host TRACEABILITY를 추가했고 e99ccecc20a72d9531d77be5b854ae6090500aff에서 신규 runner를 repository 100-column Python lint boundary에 맞췄습니다.

run_structure_noninferiority_experiment.py는 기존 owner를 조합합니다: registration validation → exact host gate → corpus admission → canonical CQT/STFT quality/performance consumer → admission/evidence identity cross-check → schema-v1 track receipts → canonical aggregation/bootstrap → result admission → atomic path-free execution envelope. MIR/admission/aggregation 로직을 복제하지 않습니다.

아직 닫히지 않은 scientific decision

Implementation 측에서는 metric/runtime identity, timing/RSS producer, track/aggregate receipt binding, exact execution-host gate, canonical end-to-end run path가 닫혔습니다. Production acceptance 전에 다음을 candidate 결과를 보지 않은 상태에서 승인·동결해야 합니다.

• rights-cleared real corpus membership과 대표성
• production noninferiority margins와 maximum candidate latency ratio
• 실제 bootstrap resample count와 random seed
• 선택한 macOS/Windows structure-performance-host-v1 payload/content address와 population/runtime claim boundary
• 해당 corpus 규모와 paired-inference plan의 충분성

Host content address는 CPU frequency/thermal state/power policy/scheduler contention/page-cache state를 완전히 고정한다고 주장하지 않습니다. 이런 host-wide variance는 scientific review와 claim boundary에 남습니다.

이 선택을 동결한 뒤 complete admitted corpus를 canonical runner로 한 번 실행합니다. 통과하더라도 production default 변경은 별도 PR입니다. Current exact head의 terminal protected checks와 independent non-author approval도 별도 merge gate입니다.

참고 기준

• Raffel, C., McFee, B., Humphrey, E. J., Salamon, J., Nieto, O., Liang, D., & Ellis, D. P. W. (2014). mir_eval: A transparent implementation of common MIR metrics. ISMIR.
• Hall, P. (1988). Theoretical comparison of bootstrap confidence intervals. The Annals of Statistics, 16(3), 927–953. https://doi.org/10.1214/aos/1176350933
• MIREX 2025 Music Structure Analysis / ismir-mirex/mirex-evaluation@b9fa0b0b32e2145af31f35830f78fc9d09a4301b.
• Python 3.14 time.perf_counter_ns(); NumPy quantile(method="linear").
• Apple sysctl, getrusage(2); Microsoft processor registry guidance and PROCESS_MEMORY_COUNTERS.PeakWorkingSetSize.

Related: #1223, #1228

[seonghobae]

#1228 Signal-MIR owner update — exact head 2b1aa6c7e230818951804f53b9425330a1bd7c54, protected base develop@314ddeae7b775a4957594b599358c8255617eb2e, ahead 127 / behind 0, Draft / mergeable.

Fresh authority review corrected the functional-ACC preregistration. The current official ismir-mirex/mirex-evaluation MIREX-2025 reproduction at b9fa0b0b32e2145af31f35830f78fc9d09a4301b uses calculate_accuracy(..., frame_hop=0.2) with np.arange(0, gt_duration, frame_hop) and t >= segment_end boundary advancement. The previous 100 ms BandScope contract was therefore not exact evaluator parity.

Current lineage:

• RED 6c20500bf31be88cd145f30fe7355a915066bdfc → GREEN 0f1e3e732154ff0ea94beb5a60e5702ed46c2453: exact upstream commit/function, 200 ms grid, frame_grid_contract_version=1.0.
• Fixtures 86e6be0b41547ecd14b4f837ff6c88cf8f770b0b / 02c73b1fff55def189bed8fe2d34d4b5ee83cb0a consume the same closed-world contract.
• ACC adapter RED 5cffef1a5db37b700ee7a27012c0f8324855665f → GREEN e72b29565a6d906a9955907656cbae97cffa471a; consolidation c84c0acdf25159f9aa066e21f9420d9fdf72c50a / fe3ed61a3b1926af6b39b6302d6f857979f9bccc reuses the parser-owned segment contract instead of duplicating it.
• TRACEABILITY is current at f03a6cedd9f8ba00dd646485323be71d417361cf / 2b1aa6c7e230818951804f53b9425330a1bd7c54.

The adapter is limited to preregistered normalized labels. Upstream raw-label normalization (solo→inst, unknown→other, etc.) is not executed after preregistration; any source mapping must happen before annotation hashing/admission.

Remaining scientific gap: wire the parser + pinned ACC adapter into the exact admitted-track consumer with the other recognized structure metrics, then implement the reviewed preregistered aggregation/paired-uncertainty procedure and execute paired CQT/STFT on the same rights-cleared real PCM/runtime identity. Production sections/segmenter.py remains unchanged on chroma_cqt.

Exact-head CI/security/native-build/SBOM/SAST runs are fresh and currently non-terminal; predecessor evidence is not counted as merge evidence.