Optimize: Tighten Heuristic thresholds and disable dithering to fix AI false negatives

app/core/audio_processor.py

@@ -166,11 +166,11 @@ class AudioProcessor:
         # 2. Normalize Volume (Re-enabled: Necessary for consistent chunked inference)
         waveform = librosa.util.normalize(waveform)
 
-        # 3. Dithering / Noise Injection
-        # Real microphones have thermal noise. High-quality digital files (AAC/AI) can be "too clean" (zeros).
-        # Adding tiny Gaussian noise prevents "perfect silence" from triggering AI detection.
-        noise_amp = 0.005 * np.max(np.abs(waveform))
-        waveform = waveform + noise_amp * np.random.normal(size=len(waveform))
+        # 3. Dithering / Noise Injection (DISABLED)
+        # Reason: Dithering added noise that helped Human audio pass, but also helped High-Quality AI pass.
+        # We now rely on the Heuristic Override to catch compressed human audio, so we don't need to fake the noise.
+        # noise_amp = 0.005 * np.max(np.abs(waveform))
+        # waveform = waveform + noise_amp * np.random.normal(size=len(waveform))
         
         return waveform
     

app/core/detector.py

@@ -316,12 +316,15 @@ class DeepfakeDetector:
             
             print(f"   features: centroid_var={centroid_var:.2f}, zcr_var={zcr_var:.4f}")
             
-            # Thresholds derived from analysis of 'SH 69 4.mp3' (Human) vs AI samples
+            # Thresholds tuned to distinguish "Noisy/Compressed Human" from "Clean/Hifi AI"
             # Human (SH 69 4): centroid_var ~774k, zcr_var ~0.011
-            # AI (ElevenLabs): often lower variance or very specific high-freq patterns
+            # AI (ElevenLabs): typically lower variance (cleaner, more consistent channel)
             
-            # If BOTH are substantial, it's likely human
-            if centroid_var > 100000 and zcr_var > 0.001:
+            # STAGE 2 TUNING:
+            # Raised zcr_var to 0.005 (was 0.001). ElevenLabs is usually < 0.005.
+            # Raised centroid_var to 250k (was 100k).
+            # This ensures only "very dynamic/noisy" recordings trigger the override.
+            if centroid_var > 250000 and zcr_var > 0.005:
                 return True
                 
             return False