Skip to content

Calibration.md

Latest commit

 

History

History
47 lines (37 loc) · 2.76 KB

Calibration.md

File metadata and controls

47 lines (37 loc) · 2.76 KB
Raw

AudioIntelligence Calibration Manifest

The official record of what the measurement layer is calibrated and verified against. Each row is backed by a test in Tests/. "Verified" means high-precision agreement with a reference on the tested scenario — not exhaustive coverage, and not the statistical estimation layer (tempo/key/instrument), which is reported separately in the README Validation Status.

🧪 Verified scenarios

Scenario Reference Result Source of truth
Integrated loudness (EBU SQAM) ffmpeg ebur128 Δ ≤ 0.08 LU (18/18) reference meter
True peak ITU-R BS.1770 Δ ≤ 0.27 dB reference meter
Loudness range (LRA) EBU Tech 3342 Δ ≤ 0.21 LU reference meter
Reference calibration (1 kHz @ −23 dBFS) EBU Tech 3341 −22.994 LUFS reference signal
Gating / silence rejection EBU Tech 3341 pass reference signal
AES17 THD+N / SMPTE IMD known-distortion signals exact synthetic references
ITU-R 468 noise weighting the standard curve ±0.03 dB analytic reference
STFT / mel librosa 0.11 corr 1.00000, 0% residual librosa parity
A-weighting (IEC 61672-1) closed-form analytic curve Δ ≤ 0.01 dB through 100Hz–2kHz analytic reference
Bit depth / sample rate / duration container header exact header read

2. Methodology

The measurement engines are verified two ways:

  • Reference signals ("digital twins") — synthetic stimuli with a known answer (e.g. a 1 kHz sine at −23 dBFS for loudness calibration), analyzed in-memory to remove I/O variance.
  • Reference implementations — ffmpeg ebur128 for loudness, librosa 0.11 for STFT/mel/MFCC parity (used only at test time; the library ships zero dependencies).

The estimation engines (tempo, key) are additionally benchmarked on real music (GiantSteps, MIREX-annotated) — there we report measured accuracy, not perfection (see README).

3. Known limitations

  • CQT engine: its correctness bugs (aliasing filter, energy rescale, octave alignment, note order) were fixed and independently cross-checked against a reference implementation (see DEVLOG Phase 10) — it is validated as a standalone engine. Key and chroma still rely on a high-resolution STFT chromagram, not CQT — but CQT does now have a real downstream consumer: TraditionalTheoryEngine.detectBassNote (chord inversion labeling, and chord root/quality tie-breaking on chroma-identical chords), wired to production's real per-chunk CQT output.
  • Neural stem separation: the NeuralSeparationEngine is an interface only — no Core ML model ships, and it is not part of analyze().

Last reviewed: 2026-09-04 — AudioIntelligence 8.2.3. See Integration.md.