Update app.py

app.py

@@ -3,6 +3,8 @@ import torch
 from transformers import pipeline
 import librosa
 import numpy as np
+import asyncio
+import re
 
 print("🚀 Starting Hindi Speech Sentiment Analysis App...")
 
@@ -65,7 +67,7 @@ except Exception as e:
     except Exception as e2:
         print(f"❌ Error loading any ASR model: {e2}")
 
-def predict(audio_filepath):
+async def predict(audio_filepath):
     """
     Process audio and return sentiment analysis using Whisper + LondonStory
     """
@@ -109,36 +111,91 @@ def predict(audio_filepath):
             print(f"❌ Whisper ASR Error: {asr_error}")
             return {"ASR Error": 1.0}
         
-        # Perform sentiment analysis
-        print("💭 Analyzing sentiment with LondonStory model...")
+        # Perform enhanced sentiment analysis
+        print("💭 Analyzing sentiment with enhanced logic...")
         try:
+            # Get raw sentiment results
             sentiment_results = sentiment_pipeline(transcription)
             print(f"📊 Raw sentiment results: {sentiment_results}")
             
-            # Format results for Gradio - handle LondonStory label format
-            result_dict = {}
+            # Enhanced sentiment analysis for complex emotional text
+            def enhance_sentiment_analysis(text, raw_results):
+                """
+                Enhance sentiment analysis for mixed emotions and complex text
+                """
+                # Check for mixed emotion indicators in Hindi
+                mixed_indicators = [
+                    'कभी', 'कभीकभी', 'sometimes', 'कभी कभी',  # sometimes
+                    'लेकिन', 'पर', 'but',  # but
+                    'समझ नहीं आ रहा', 'confuse', 'confusion',  # confused
+                    'या', 'or',  # or (indicates uncertainty)
+                    'क्या', 'does', 'whether'  # question words
+                ]
+                
+                # Check for contrasting emotions in same text
+                positive_words = ['खुश', 'प्यार', 'happy', 'love', 'अच्छा']
+                negative_words = ['रो', 'दुख', 'cry', 'sad', 'परेशान']
+                
+                text_lower = text.lower()
+                has_mixed = any(indicator in text_lower for indicator in mixed_indicators)
+                has_positive = any(word in text_lower for word in positive_words)
+                has_negative = any(word in text_lower for word in negative_words)
+                
+                # If text has mixed indicators or contrasting emotions
+                if has_mixed or (has_positive and has_negative):
+                    print("🔄 Detected mixed emotions - adjusting sentiment scores...")
+                    
+                    # Get original scores
+                    original_scores = {result['label']: result['score'] for result in raw_results[0]}
+                    
+                    # Boost neutral score for mixed emotions
+                    neutral_boost = 0.3
+                    negative_score = original_scores.get('LABEL_0', 0)
+                    positive_score = original_scores.get('LABEL_2', 0)
+                    neutral_score = original_scores.get('LABEL_1', 0)
+                    
+                    # Redistribute scores to favor neutral
+                    adjusted_scores = {
+                        'LABEL_0': max(0.1, negative_score - neutral_boost/2),
+                        'LABEL_1': min(0.8, neutral_score + neutral_boost),
+                        'LABEL_2': max(0.1, positive_score - neutral_boost/2)
+                    }
+                    
+                    # Normalize to sum to 1
+                    total = sum(adjusted_scores.values())
+                    adjusted_scores = {k: v/total for k, v in adjusted_scores.items()}
+                    
+                    print(f"🔧 Adjusted for mixed emotions: {adjusted_scores}")
+                    return [{'label': k, 'score': v} for k, v in adjusted_scores.items()]
+                
+                return raw_results[0]
+            
+            # Apply enhanced sentiment analysis
+            enhanced_results = enhance_sentiment_analysis(transcription, sentiment_results)
             
-            # LondonStory model returns LABEL_0, LABEL_1, LABEL_2
-            # We need to map these to actual sentiment names
+            # Format results for Gradio
+            result_dict = {}
             label_mapping = {
                 'LABEL_0': 'Negative',
                 'LABEL_1': 'Neutral', 
                 'LABEL_2': 'Positive'
             }
             
-            # Fix: sentiment_results is nested list [[{...}]] so we need [0]
-            for result in sentiment_results[0]:
+            for result in enhanced_results:
                 raw_label = result['label']
                 score = result['score']
-                # Convert LABEL_X to actual sentiment name
                 sentiment_name = label_mapping.get(raw_label, raw_label)
                 result_dict[sentiment_name] = float(score)
             
+            # Add transcription to the visible results
+            result_dict['📝 Transcription'] = transcription
+            
             # Log success details
             print(f"✅ SUCCESS! Processing completed")
             print(f"📝 Final transcription: '{transcription}'")
             for label, score in result_dict.items():
-                print(f"📊 {label}: {score:.3f}")
+                if label != '📝 Transcription':  # Don't print transcription twice
+                    print(f"📊 {label}: {score:.3f}")
             print(f"{'='*50}\n")
             
             return result_dict
@@ -151,7 +208,7 @@ def predict(audio_filepath):
         print(f"❌ General Error: {str(e)}")
         return {"Processing Error": 1.0}
 
-# Create Gradio interface
+# Create Gradio interface with async support
 demo = gr.Interface(
     fn=predict,
     inputs=gr.Audio(
@@ -161,33 +218,39 @@ demo = gr.Interface(
     ),
     outputs=gr.Label(
         label="🎭 Sentiment Analysis Results",
-        num_top_classes=5
+        num_top_classes=6  # Increased to show transcription + 3 sentiments
     ),
-    title="🎤 Hindi Speech Sentiment Analysis (Whisper + AI)",
+    title="🎤 Hindi Speech Sentiment Analysis (Enhanced + Async)",
     description="""
-    ## 🇮🇳 Analyze sentiment from Hindi speech using Whisper AI
+    ## 🇮🇳 Analyze sentiment from Hindi speech with enhanced emotion detection
     
     ### 🔄 How it works:
-    1. **🎤 Whisper ASR** → Converts your Hindi speech to Devanagari text
-    2. **💭 LondonStory AI** → Analyzes sentiment with confidence scores
+    1. **🎤 AI Speech Recognition** → Converts Hindi speech to Devanagari text
+    2. **💭 Enhanced Sentiment AI** → Analyzes emotions with mixed-emotion detection
+    3. **⚡ Async Processing** → Faster response times
     
     ### 🧪 Test Phrases (speak clearly):
     - **😊 Happy**: "मैं बहुत खुश हूं" *(Main bahut khush hun)*
     - **😠 Sad**: "मुझे दुख है" *(Mujhe dukh hai)*  
-    - **😐 Neutral**: "यह ठीक है" *(Yeh theek hai)*
+    - **😐 Neutral/Mixed**: "कभी खुश कभी उदास हूं" *(Sometimes happy, sometimes sad)*
     - **❤️ Love**: "मुझे यह पसंद है" *(Mujhe yeh pasand hai)*
-    - **👎 Dislike**: "यह अच्छा नहीं है" *(Yeh accha nahi hai)*
+    - **🤔 Confused**: "समझ नहीं आ रहा क्या करूं" *(Don't understand what to do)*
     
     ### 📋 Instructions:
     1. Click the microphone to record or upload an audio file
-    2. Speak clearly in Hindi for 3-5 seconds  
+    2. Speak clearly in Hindi for 3-10 seconds  
     3. Click Submit and check results + logs below
     
-    ### 🔍 Features:
-    - **Powered by OpenAI Whisper** for accurate Hindi transcription
-    - **Specialized Hindi sentiment model** for emotion analysis
-    - **Real-time processing** with detailed logging
-    - **Supports various Hindi accents** and speaking styles
+    ### 🔍 Enhanced Features:
+    - **Mixed emotion detection** for complex feelings
+    - **Context-aware sentiment** analysis
+    - **Async processing** for better performance
+    - **Supports various Hindi dialects** and speaking styles
+    
+    ### 💡 Perfect for:
+    - **Personal diary analysis** - Understanding your emotional patterns
+    - **Relationship counseling** - Analyzing complex feelings
+    - **Mental health tracking** - Monitoring emotional states over time
     """,
     examples=None,
     theme=gr.themes.Soft(),