@newgameplusinc/odyssey-audio-video-sdk-dev 1.0.352 → 1.0.353

package/dist/audio/MLNoiseSuppressor.js

@@ -1122,22 +1122,24 @@ class MLNoiseSuppressor {
                 gains[k] = filtTotal[k] > 1e-8 ? gains[k] / filtTotal[k] : 1.0;
                 gains[k] = Math.max(0.0, Math.min(1.0, gains[k]));
             }
-            // ── Step 5: Wiener-style gain compression ─────────────────────────────
-            // Problem: fan noise is in the same freq bands as speech (100-500 Hz).
-            // The model gives IRM≈0.85 to speech bins AND to fan bins that overlap,
-            // so the fan rides through at 85% amplitude on speech-containing bins.
+            // ── Step 5: Speech-band protection floor ──────────────────────────────
+            // β=1.8 power law was crushing voice to 12% amplitude (mean IRM=0.12
+            // across 257 bins) — user heard noise only, no voice. Root cause:
+            // β^1.8 was computed before STFT was applied (binary gate hid the damage);
+            // now that per-bin STFT gains are real, voice must be protected.
             //
-            // Fix: apply power law  gains[k] = gains[k]^β  (β > 1).
-            // This is mathematically equivalent to Wiener spectral subtraction and
-            // is used by every production noise suppressor (RNNoise, Google, WebRTC).
+            // During speech frames: enforce minimum gain of 0.20 on all bins.
             // Effect:
-            //   IRM=0.85 (strong speech) → 0.85^1.8 = 0.763  (−2.3dB, speech fine)
-            //   IRM=0.60 (mixed)         → 0.60^1.8 = 0.425  (−7.4dB, fan partly removed)
-            //   IRM=0.35 (fan+speech)    → 0.35^1.8 = 0.143  (−16.9dB, fan mostly gone)
-            //   IRM=0.10 (pure noise)    → 0.10^1.8 = 0.016  (−36dB, near silence)
-            const GAIN_BETA = 1.8;
-            for (let k = 1; k < bins; k++) {
-                gains[k] = Math.pow(gains[k], GAIN_BETA);
+            //   Speech fundamental (raw IRM≈0.86) → 0.86  (unchanged, full voice)
+            //   Speech formants    (raw IRM≈0.65) → 0.65  (unchanged, full voice)
+            //   Mixed fan+speech   (raw IRM≈0.30) → 0.30  (−10dB vs binary gate)
+            //   Pure fan bins      (raw IRM≈0.10) → 0.20  (−14dB vs binary gate)
+            // During noise frames: no floor — model can suppress to full silence.
+            const IRM_SPEECH_FLOOR = 0.2;
+            if (isSpeechFrame) {
+                for (let k = 1; k < bins; k++) {
+                    gains[k] = Math.max(IRM_SPEECH_FLOOR, gains[k]);
+                }
             }
             // ── Step 6: Override gains[0] as explicit gate signal ────────────────
             // The DC bin (k=0) has no mel filter coverage → filtTotal[0] ≤ 1e-8 →

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@newgameplusinc/odyssey-audio-video-sdk-dev",
-  "version": "1.0.352",
+  "version": "1.0.353",
   "description": "Odyssey Spatial Audio & Video SDK using MediaSoup for real-time communication",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",