@audio/denoise-core 0.1.0 → 0.1.2

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@audio/denoise-core",
-  "version": "0.1.0",
-  "description": "Shared restoration primitives — STFT (batch/stream/analyse), noise estimation (min-stats, IMCRA), VAD, AR modeling, quality metrics",
+  "version": "0.1.2",
+  "description": "Internal glue for @audio/denoise-* atoms — STFT (batch/stream/analyse), noise estimation (min-stats, IMCRA), AR modeling, quality metrics. VAD lives in @audio/vad (re-exported here).",
   "type": "module",
   "sideEffects": false,
   "main": "index.js",
@@ -25,7 +25,8 @@
     "quality.js"
   ],
   "dependencies": {
-    "fourier-transform": "^2.0.0"
+    "fourier-transform": "^2.0.0",
+    "@audio/vad": "^1.0.0"
   },
   "keywords": [
     "audio",

package/stft.js

@@ -93,6 +93,7 @@ export function stftStream(process, opts) {
         ob[base + i] += sf[i] * win[i]
         nb[base + i] += win[i] * win[i]
       }
+      st.hi = Math.max(st.hi, st.pos + N)
       aPos += hop
       st.pos += hop
     }

package/util.js

@@ -120,7 +120,7 @@ export function makeStreamBufs(N, nf = 0) {
     N, nf,
     ib: new Float32Array(N * 4), il: 0,
     ob: new Float32Array(N * 8), nb: new Float32Array(N * 8),
-    pos: 0, oread: 0
+    pos: 0, oread: 0, hi: 0
   }
 }
 
@@ -156,9 +156,11 @@ export function take(st, upTo) {
   }
   st.oread += len
   if (st.oread > st.N * 8) {
+    // shift left; zero only past the high-water mark — frames extend N−hop beyond
+    // pos, so zeroing from pos would erase the last frame's partial overlap-add tail
     st.ob.copyWithin(0, st.oread); st.nb.copyWithin(0, st.oread)
-    st.pos -= st.oread; st.oread = 0
-    st.ob.fill(0, st.pos); st.nb.fill(0, st.pos)
+    st.pos -= st.oread; st.hi = Math.max(0, st.hi - st.oread); st.oread = 0
+    st.ob.fill(0, st.hi); st.nb.fill(0, st.hi)
   }
   return out
 }

package/vad.js

@@ -1,91 +1,3 @@
-// Voice Activity Detection + Speech Presence Probability.
-//   - vad: hard 0/1 decision per frame from energy + spectral flatness + ZCR
-//   - spp: soft probability per bin from a-priori SNR (drives OM-LSA, IMCRA)
-//
-// Use vad() for gating decisions (debreath, gate); spp() for spectral denoise gain.
-
-import { stftAnalyse } from './stft.js'
-import { lin2db } from './util.js'
-
-// Returns { active: Uint8Array(frames), times: Float32Array(frames) of frame-start sec }.
-// Active iff energy is N dB above the rolling noise floor AND spectral flatness < flatTh.
-//
-// Floor estimation: minimum of energy over a sliding D-frame window with bias
-// correction — Martin-style minimum tracking. This avoids the "speech eats its own
-// floor" failure mode of plain exponential smoothing.
-//
-// flatTh ≈ 0.4 — speech is tonal (low flatness), noise is flat (high flatness).
-export function vad(data, opts = {}) {
-  let N = opts.frameSize || 1024
-  let hop = opts.hopSize || (N >> 1)
-  let fs = opts.fs || 44100
-  let snrTh = opts.snrTh ?? 6                  // dB above floor
-  let flatTh = opts.flatTh ?? 0.4              // spectral flatness threshold
-  let D = opts.window || 64                    // frames in min-tracker window
-  let bias = opts.bias ?? 5                    // dB — added to min to estimate true floor
-
-  let frames = Math.max(0, Math.floor((data.length - N) / hop) + 1)
-  let active = new Uint8Array(frames)
-  let times = new Float32Array(frames)
-  let energies = new Float64Array(frames)
-  let flats = new Float64Array(frames)
-  let i = 0
-
-  stftAnalyse(data, (mag, _phase, pos) => {
-    let half = mag.length - 1, e = 0
-    let logSum = 0, linSum = 0, nz = 0
-    for (let k = 1; k <= half; k++) {
-      let p = mag[k] * mag[k]
-      e += p
-      if (p > 1e-30) { logSum += Math.log(p); nz++ }
-      linSum += p
-    }
-    let geom = nz ? Math.exp(logSum / nz) : 0
-    let arith = linSum / Math.max(half, 1)
-    flats[i] = arith > 1e-30 ? geom / arith : 1
-    energies[i] = lin2db(Math.sqrt(e / N))
-    times[i] = pos / fs
-    i++
-  }, { frameSize: N, hopSize: hop })
-
-  // Global noise floor = 10th-percentile energy + bias. Robust on signals where
-  // any short window may be entirely speech.
-  let sorted = Float64Array.from(energies).sort()
-  let floor = sorted[Math.floor(sorted.length * 0.1)] + bias
-
-  for (let j = 0; j < frames; j++) {
-    active[j] = (energies[j] - floor > snrTh && flats[j] < flatTh) ? 1 : 0
-  }
-  return { active, times, hop, frameSize: N }
-}
-
-// Per-bin Speech Presence Probability from a-priori SNR ξ:
-//   p = ξ / (1 + ξ)   (Bayesian formulation under Gaussian model, q-prior = 0.5)
-// Bind to a noise PSD source (e.g. minStats.psd or imcra.psd) and an a-priori SNR estimate.
-export function spp(mag, noisePsd, opts = {}) {
-  let xiMin = opts.xiMin ?? 0.0316             // -15 dB floor
-  let half = mag.length - 1
-  let p = new Float64Array(half + 1)
-  for (let k = 0; k <= half; k++) {
-    let post = (mag[k] * mag[k]) / Math.max(noisePsd[k], 1e-30) - 1
-    let xi = Math.max(xiMin, post)
-    p[k] = xi / (1 + xi)
-  }
-  return p
-}
-
-// Decision-Directed a-priori SNR (Ephraim-Malah 1984), recursive smoothing.
-//   ξ̂[k] = α · |X̂_prev[k]|² / N̂[k]  +  (1-α) · max(γ-1, 0)
-// γ = posterior SNR = |Y[k]|² / N̂[k]. Used by Wiener, MMSE, OM-LSA gain rules.
-export function ddSnr(mag, noisePsd, prevGain, prevMag, alpha = 0.98) {
-  let half = mag.length - 1
-  let xi = new Float64Array(half + 1)
-  for (let k = 0; k <= half; k++) {
-    let n = Math.max(noisePsd[k], 1e-30)
-    let post = (mag[k] * mag[k]) / n
-    let g = prevGain[k] * prevMag[k]
-    let prev = (g * g) / n
-    xi[k] = Math.max(0.0316, alpha * prev + (1 - alpha) * Math.max(post - 1, 0))
-  }
-  return xi
-}
+// vad/spp/ddSnr were promoted to the standalone @audio/vad atom.
+// Re-exported here so restoration code (and the @audio/denoise umbrella) keep working.
+export * from '@audio/vad'