@newgameplusinc/odyssey-audio-video-sdk-dev 1.0.341 → 1.0.342

package/dist/audio/MLNoiseSuppressor.js

@@ -209,21 +209,17 @@ tf.serialization.registerClass(GRUCellResetAfterSupport);
 //
 // WEIGHT NAMING:
 //   model.json weightsManifest uses paths like  "gru_96/gru_cell/kernel".
-//   TF.js builds variable names as  `${rnnLayer.name}/${cell.name}/${varName}`.
-//   The RNN wrapper contributes "${layerName}/" automatically, so the cell
-//   must be named just "gru_cell" (no layer prefix).
-//   Result: "gru_96" + "/" + "gru_cell" + "/" + "kernel" = "gru_96/gru_cell/kernel" ✅
-//   If cell were named "${layerName}/gru_cell" (old, wrong):
-//   "gru_96" + "/" + "gru_96/gru_cell" + "/" + "kernel" = "gru_96/gru_96/gru_cell/kernel" ❌
+//   TF.js builds variable names as  `${layer.name}/${varName}`.
+//   So the cell MUST be named `${layerName}/gru_cell`  (e.g. "gru_96/gru_cell")
+//   so that addWeight('kernel') → "gru_96/gru_cell/kernel"  ← matches manifest.
 // =============================================================================
 class GRULayerWithResetAfter {
     static fromConfig(_cls, config) {
         const layerName = config.name;
-        // Cell named "gru_cell" only — RNN wrapper prepends layerName automatically.
-        // Final weight path: "${layerName}/gru_cell/kernel" ✅
+        // Create our cell — named so its weights match the manifest paths
         const cell = new GRUCellResetAfterSupport({
             ...config,
-            name: `gru_cell`, // RNN layer adds "gru_96/" prefix → "gru_96/gru_cell/kernel" ✅
+            name: `${layerName}/gru_cell`, // → weights: "gru_96/gru_cell/kernel" ✅
             // Normalise keys: Keras JSON is snake_case; TF.js internals are camelCase
             useBias: config.useBias ?? config.use_bias ?? true,
             recurrentActivation: config.recurrentActivation ?? config.recurrent_activation ?? "sigmoid",
@@ -351,21 +347,13 @@ class MLNoiseSuppressor {
             const nMels = this.config.n_mels || 40;
             console.log(`[MLNoiseSuppressor] Warming up model (${seqLen} × ${nMels})...`);
             const warmupInput = tf.zeros([1, seqLen, nMels]);
-            let warmupMax = 0;
             for (let w = 0; w < 3; w++) {
                 const warmupOut = this.model.predict(warmupInput);
-                const warmupData = warmupOut.dataSync();
-                for (let i = 0; i < warmupData.length; i++)
-                    if (warmupData[i] > warmupMax)
-                        warmupMax = warmupData[i];
+                warmupOut.dataSync(); // force full synchronous execution
                 warmupOut.dispose();
             }
             warmupInput.dispose();
-            // For zero input, sigmoid(bias_only) should produce non-zero output if weights loaded.
-            // If warmupMax === 0, weights failed to load (name mismatch in weightsManifest).
-            console.log(`[MLNoiseSuppressor] Warmup done — zero-input output max: ${warmupMax.toFixed(4)} ${warmupMax > 0
-                ? "✅ weights loaded"
-                : "❌ weights NOT loaded — check weight name mapping"}`);
+            console.log(`[MLNoiseSuppressor] Warmup done`);
             this.isInitialized = true;
             console.log(`[MLNoiseSuppressor] ✅ Ready — noise suppression is ACTIVE`);
             console.log(`[MLNoiseSuppressor] Config: ${modelSampleRate}Hz, ${this.config.n_mels} mels, n_fft=${this.config.n_fft || 2048}`);
@@ -1019,13 +1007,6 @@ class MLNoiseSuppressor {
                 if (rawMask[m] > rawMax)
                     rawMax = rawMask[m];
             }
-            // Model failure fallback: if model outputs all-zeros (wrong/cached model files),
-            // returning gains=0 suppresses all speech. Return full passthrough instead.
-            if (rawMax === 0) {
-                const passthrough = new Float32Array(bins).fill(1.0);
-                passthrough[0] = 1.0; // speech flag = on, keeps worklet gate open
-                return passthrough;
-            }
             // Threshold 0.108: fan/AC noise consistently scores rawMax=0.094-0.100.
             // Speech consistently scores rawMax≥0.111. Gap between 0.100 and 0.111 gives
             // 8% margin. Previously 0.10 caused fan frames scoring exactly 0.100 to
@@ -1084,16 +1065,6 @@ class MLNoiseSuppressor {
                     gains[k] = Math.max(IRM_SPEECH_FLOOR, gains[k]);
                 }
             }
-            // ── Step 6: Set DC bin as explicit speech/noise gate signal ──────────
-            // The worklet reads gains[0] to control its ML gate:
-            //   gains[0] >= 0.5  → speech detected → open gate + reset 533ms holdover
-            //   gains[0] <  0.5  → noise frame    → count down holdover, then close
-            //
-            // DC bin (0 Hz) has no mel filter coverage so it always computes to 1.0
-            // above (no-coverage default). This means the ML gate NEVER closes via ML
-            // — fan noise leaks through during inter-word pauses. Override it explicitly
-            // so the worklet gate actually follows the ML speech decision.
-            gains[0] = isSpeechFrame ? 1.0 : 0.0;
             // ── Diagnostic log ────────────────────────────────────────────────────
             const now = Date.now();
             const lastLog = MLNoiseSuppressor._lastGainLog || new Map();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@newgameplusinc/odyssey-audio-video-sdk-dev",
-  "version": "1.0.341",
+  "version": "1.0.342",
   "description": "Odyssey Spatial Audio & Video SDK using MediaSoup for real-time communication",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",