@audio/defeedback-analyzer

Feedback candidate detection — spectral peaks scored by PNPR/PHPR howl criteria (van Waterschoot & Moonen 2011)

import analyze from '@audio/defeedback-analyzer'

Windows one analysis frame, FFTs it, and picks spectral peaks scored by two classical howl criteria: PNPR (peak-to-neighboring-power ratio — feedback is far narrower than any musical content) and PHPR (peak-to-harmonic-power ratio — a bare howl carries no 2nd/3rd harmonic, a musical tone does). Peak frequencies are parabolically refined in the dB domain (~0.1% accuracy). A relational harmonic gate then drops any peak that sits within 1% of an integer multiple of a stronger lower peak — a musical overtone, not a second howl.

let candidates = analyze(frame, { fs: 48000 })
// [{ freq, level (dB), pnpr, phpr, kc }, ...] — sorted by level, strongest first
// kc: fractional FFT bin index of the peak

data is one power-of-2 analysis window (Hann-windowed internally); bins below 100 Hz are ignored (rumble). Feed candidates into @audio/defeedback-tracker for persistence gating before deploying a notch.

Use when: building a custom AFS pipeline stage-by-stage.
Not for: end-to-end suppression — use the @audio/defeedback umbrella factory.

Part of @audio/defeedback — the defeedback family umbrella.

MIT © audiojs