@audio/spectral

Spectral features — centroid, spread, flatness, rolloff, flux, slope, crest, zcr. Shipped.

Also shipped: spectral-mfcc (librosa parity), spectral-ltas (adaptive/match-EQ substrate), spectral-edit (time×frequency region editing — Audacity/afftfilt class), spectral-harmonics (inharmonicity/tristimulus/odd-even from f0).

Parity: FFmpeg aspectralstats, MIREX spectral features. LTAS / spectral-target power the Auto-Chain analyser's adaptive EQ.

spectral-target

Target-curve library — per-content-type LTAS targets ('speech', 'music', 'pink', 'voice-music', or a custom anchor table) plus deviation(), the math that turns a measured spectral-ltas curve and a target into an EQ correction curve. The adaptive-EQ data+math layer behind Mix Analyser's spectral-balance report, Speech/Music Enhancer's adaptive EQ, Auto-Chain's Stage-2 EQ, and Match-by-Reference.

• 'speech' — Byrne, Dillon, Tran et al. 1994, JASA 96(4):2108–2120, Table II "Combined" column (12-language universal LTASS, 100 Hz–10 kHz).
• 'music' — Pestana, Ma, Reiss, Barbosa & Black 2013, AES Convention 135 Paper 8960, p.5: "a linearly decaying distribution of around 5 dB per octave between 100 and 4000 Hz."
• 'pink' — analytic −3.0103 dB/octave, no citation needed.
• 'voice-music' — audiojs-defined convention, not literature: bin-wise max(speech, music) − 6 dB.

Edge policies: below the lowest anchor holds flat, above the highest continues the last segment's slope, interpolation is log-frequency linear. Every curve is mean-normalized to 0 dB over its own defined band; deviation() band-limits and tapers to exactly 0 outside [20 Hz, 0.45·fs]. Full citations, exact tables, and the manifest-less rationale are in packages/spectral-target/README.md.