Instrument Identifier AI

This is a signal-analysis identifier, not a magic black box. Every detected note event is measured the way timbre research measures it: MPEG-7-style log-attack-time (time from 2% to 100% of peak energy, Peeters 2000), harmonic power ratio, the odd/even partial power ratio over harmonics 1-5 that separates a clarinet’s cylindrical bore from a trumpet’s flare (Almeida 2024), the string-stretch inharmonicity coefficient B from fₙ = n·f₀·√(1+B·n²) that fingerprints piano vs guitar strings (Fletcher 2000), spectral centroid/spread/flatness, and a 4-7 Hz vibrato detector that is the classic vocal/violin signature. Each event is scored against 13 profiles (piano, guitar, bass, violin, flute, clarinet, sax, brass, organ, synth lead/pad, voice, drums) and rolled up into instrument families. Be realistic about polyphonic mixes: state-of-the-art classifiers reach roughly 35% per-instrument and 77% per-family accuracy on isolated notes (Eronen 2001), so treat results as a strong hint, strongest on exposed solos, stems and one-instrument loops.