Upload pipeline.py with huggingface_hub

pipeline.py

@@ -1,8 +1,3 @@
-"""
-Multimodal Authenticity Detection Pipeline
-Integrates CNN audio classification, Whisper ASR, and text authenticity analysis
-"""
-
 from typing import Dict, Optional
 import time
 from audio_classifier import AudioClassifier
@@ -102,22 +97,32 @@ class AuthenticityDetectionPipeline:
         text_results: Dict
     ) -> Dict:
 
+        # CNN score: spontaneous = authentic (high), read = inauthentic (low)
         if audio_results['classification'] == 'spontaneous':
             audio_score = audio_results['confidence']
         else:  # read
             audio_score = 1.0 - audio_results['confidence']
         
-        if asr_results['kopparapu_classification'] == 'spontaneous':
-            speech_pattern_score = asr_results['kopparapu_score']
-        else:
-            speech_pattern_score = 1.0 - asr_results['kopparapu_score']
+        # Kopparapu score: 0=spontaneous, 1=read
+        # Invert so spontaneous (low kopparapu) = high authenticity
+        speech_pattern_score = 1.0 - asr_results['kopparapu_score']
         
-        text_auth_score = text_results['authenticity_score']
+        # Filler words: higher ratio = more spontaneous = more authentic
+        filler_ratio = asr_results['filler_words']['ratio']
+        filler_score = min(1.0, filler_ratio / 0.05)  # Normalize: 5%+ = max score
         
+        # Pause variability: higher = more spontaneous = more authentic
+        pause_var = asr_results['pause_patterns']['pause_variability']
+        pause_score = min(1.0, pause_var / 0.5)  # Normalize: 0.5+ = max score
+        
+        text_auth_score = text_results['authenticity_score']
+
         composite_score = (
-            audio_score * 0.30 +            # CNN acoustic analysis
-            speech_pattern_score * 0.30 +   # Speech patterns (Kopparapu)
-            text_auth_score * 0.40          # Text authenticity (AI detection)
+            audio_score * 0.15 +            # CNN - weakest component
+            speech_pattern_score * 0.20 +   # Kopparapu linguistic
+            filler_score * 0.10 +           # Filler word ratio
+            pause_score * 0.05 +            # Pause variability
+            text_auth_score * 0.50          # Text authenticity - strongest signal
         )
         
         if composite_score >= 0.7:
@@ -186,4 +191,3 @@ if __name__ == "__main__":
         whisper_model_size="base"
     )
     print("\nPipeline ready for audio analysis.")
-