@audio/defeedback-analyzer 1.0.0

package/analyzer.js

@@ -0,0 +1,59 @@
+// Feedback candidate detection — spectral peaks scored by the classical howl criteria
+// (van Waterschoot & Moonen 2011): PNPR (peak-to-neighboring-power ratio — feedback is
+// far narrower than music) and PHPR (peak-to-harmonic-power ratio — feedback is a bare
+// sinusoid, musical tones carry harmonics).
+
+import { fft } from 'fourier-transform'
+
+/**
+ * Analyze one window. @returns Array<{freq, level (dB), pnpr, phpr}>
+ * @param {Float32Array} data — analysis window (power of 2)
+ * @param {object} opts — { fs=44100, maxCandidates=6, floor=-70 (dB) }
+ */
+export default function analyze (data, { fs = 44100, maxCandidates = 6, floor = -70 } = {}) {
+	let n = data.length, half = n / 2
+	let buf = new Float64Array(n)
+	for (let i = 0; i < n; i++) buf[i] = data[i] * (0.5 - 0.5 * Math.cos(2 * Math.PI * i / n))
+	let [re, im] = fft(buf)
+	let mag = new Float64Array(half + 1)
+	for (let k = 0; k <= half; k++) mag[k] = re[k] * re[k] + im[k] * im[k]
+
+	let db = p => 10 * Math.log10(p + 1e-20)
+	let at = kc => { // max power around a (possibly fractional) bin
+		let b = Math.round(kc)
+		let m = 0
+		for (let j = Math.max(0, b - 1); j <= Math.min(half, b + 1); j++) if (mag[j] > m) m = mag[j]
+		return m
+	}
+	let out = []
+	let kMin = Math.max(4, Math.round(100 * n / fs)) // ignore <100 Hz rumble
+	for (let k = kMin; k < half - 4; k++) {
+		if (mag[k] <= mag[k - 1] || mag[k] < mag[k + 1]) continue
+		let level = db(mag[k] * 4 / (n * n)) // rough dBFS calibration
+		if (level < floor) continue
+		// parabolic freq refinement (dB domain)
+		let la = Math.log(mag[k - 1] + 1e-20), lb = Math.log(mag[k] + 1e-20), lc = Math.log(mag[k + 1] + 1e-20)
+		let off = 0.5 * (la - lc) / (la - 2 * lb + lc || 1e-12)
+		if (!(off > -1 && off < 1)) off = 0
+		let kc = k + off
+		// PNPR: peak vs mean power a few bins away (±4..8) — feedback is far narrower than music
+		let nb = (mag[k - 4] + mag[k - 5] + mag[k - 6] + mag[k + 4] + mag[k + 5] + mag[k + 6]) / 6
+		let pnpr = db(mag[k]) - db(nb)
+		// PHPR: peak vs its own 2nd/3rd harmonics — a bare howl has none
+		let phpr = db(mag[k]) - db(Math.max(2 * kc <= half ? at(2 * kc) : 0, 3 * kc <= half ? at(3 * kc) : 0))
+		out.push({ freq: kc * fs / n, level, pnpr, phpr, kc })
+	}
+	// relational harmonic gate: a peak that sits within 1% of an integer multiple of a
+	// strong lower peak is a musical overtone, not feedback (parabolic freqs ≈ 0.1% accurate)
+	let fundamentals = out.filter(c => c.level > -45).sort((a, b) => a.kc - b.kc).slice(0, 8)
+	out = out.filter(c => {
+		for (let f of fundamentals) {
+			if (f.kc >= c.kc * 0.75) break
+			let m = Math.round(c.kc / f.kc)
+			if (m >= 2 && Math.abs(c.kc - m * f.kc) / (m * f.kc) < 0.01) return false
+		}
+		return true
+	})
+	out.sort((a, b) => b.level - a.level)
+	return out.slice(0, maxCandidates)
+}

package/package.json

@@ -0,0 +1,33 @@
+{
+  "name": "@audio/defeedback-analyzer",
+  "version": "1.0.0",
+  "description": "Feedback candidate detection — spectral peaks with PNPR/PHPR howl criteria (Waterschoot & Moonen 2011)",
+  "type": "module",
+  "sideEffects": false,
+  "main": "analyzer.js",
+  "exports": {
+    ".": "./analyzer.js",
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "analyzer.js"
+  ],
+  "dependencies": {
+    "fourier-transform": "^2.2.0"
+  },
+  "keywords": [
+    "audio",
+    "dsp",
+    "feedback",
+    "howling",
+    "analyzer"
+  ],
+  "license": "MIT",
+  "author": "Dmitry Iv. <dfcreative@gmail.com>",
+  "engines": {
+    "node": ">=18"
+  },
+  "publishConfig": {
+    "access": "public"
+  }
+}