Update app.py

app.py

@@ -398,15 +398,45 @@ def detect_word_boundary_overlap(audio_segment: torch.Tensor, sample_rate: int,
     if len(energy_levels) < 3:
         return 0.0
     
-    # Look for pattern: [high energy] → [low energy] → [high energy]
-    silence_threshold = np.percentile(energy_levels, 20)  # Bottom 20%
-    noise_threshold = silence_threshold * 3
+    # FIXED: Use absolute thresholds that work with real audio levels
+    # Convert to logarithmic scale to better handle wide dynamic range
+    log_energies = []
+    for energy in energy_levels:
+        # Convert to dB-like scale, with floor to prevent log(0)
+        log_energy = 10 * np.log10(max(energy, 1e-10))
+        log_energies.append(log_energy)
+    
+    # Calculate thresholds in log domain
+    min_log = min(log_energies)
+    max_log = max(log_energies)
+    dynamic_range = max_log - min_log
+    
+    # If dynamic range is too small, everything is similar energy - no clear pattern
+    if dynamic_range < 6:  # Less than 6dB difference
+        log(f"🔍 Boundary analysis for '{word}': insufficient dynamic range ({dynamic_range:.1f}dB)")
+        return 0.0
+    
+    # Set thresholds: silence is bottom 25% of range, noise is top 50%
+    silence_threshold_log = min_log + dynamic_range * 0.25
+    noise_threshold_log = min_log + dynamic_range * 0.5
     
-    log(f"🔍 Boundary analysis for '{word}': {len(energy_levels)} windows, silence_thresh={silence_threshold:.6f}, noise_thresh={noise_threshold:.6f}")
+    # Convert back to linear for comparison
+    silence_threshold = 10 ** (silence_threshold_log / 10)
+    noise_threshold = 10 ** (noise_threshold_log / 10)
     
-    # Log each energy window for detailed analysis
+    log(f"🔍 Boundary analysis for '{word}': {len(energy_levels)} windows")
+    log(f"   Energy range: {min(energy_levels):.6f} to {max(energy_levels):.6f}")
+    log(f"   Log range: {min_log:.1f}dB to {max_log:.1f}dB (dynamic range: {dynamic_range:.1f}dB)")
+    log(f"   Thresholds: silence={silence_threshold:.6f}, noise={noise_threshold:.6f}")
+    
+    # Classify windows with new thresholds
     for i, energy in enumerate(energy_levels):
-        window_type = "NOISE" if energy > noise_threshold else ("SILENCE" if energy < silence_threshold else "MID")
+        if energy > noise_threshold:
+            window_type = "NOISE"
+        elif energy < silence_threshold:
+            window_type = "SILENCE"
+        else:
+            window_type = "MID"
         log(f"   Window {i}: energy={energy:.6f} [{window_type}]")
     
     # Find sustained silence (2+ consecutive low-energy windows)