Update app.py

app.py

@@ -1,4 +1,4 @@
-import gradio as gr
+ import gradio as gr
 import torch
 from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer
 import librosa
@@ -15,19 +15,17 @@ print("🚀 Starting Enhanced Hindi Speech Sentiment Analysis App...")
 # 1. LOAD MODELS
 # ============================================
 
-# Load XLM-RoBERTa Hindi Sentiment Model (Better accuracy)
-print("📚 Loading XLM-RoBERTa sentiment analysis model...")
+# Load Hindi Sentiment Model
+print("📚 Loading Hindi sentiment analysis model...")
 try:
-    sentiment_model_name = "cardiffnlp/twitter-xlm-roberta-base-sentiment"
-    sentiment_tokenizer = AutoTokenizer.from_pretrained(sentiment_model_name)
-    sentiment_model = AutoModelForSequenceClassification.from_pretrained(sentiment_model_name)
+    # Use LondonStory's Hindi sentiment model
+    sentiment_model_name = "LondonStory/txlm-roberta-hindi-sentiment"
     sentiment_pipeline = pipeline(
         "text-classification",
-        model=sentiment_model,
-        tokenizer=sentiment_tokenizer,
+        model=sentiment_model_name,
         top_k=None
     )
-    print("✅ XLM-RoBERTa sentiment model loaded successfully")
+    print("✅ Hindi sentiment model loaded successfully")
 except Exception as e:
     print(f"❌ Error loading sentiment model: {e}")
     raise
@@ -265,21 +263,29 @@ def enhanced_sentiment_analysis(text, prosodic_features, raw_results):
     """
     Enhanced sentiment analysis combining text and prosodic features
     """
-    # Parse raw results
+    # Parse raw results - handle different model formats
     sentiment_scores = {}
+    
+    # Check if results are in the expected format
+    if not raw_results or not isinstance(raw_results, list) or len(raw_results) == 0:
+        print("⚠️ Unexpected sentiment results format")
+        return {'Negative': 0.33, 'Neutral': 0.34, 'Positive': 0.33}, 0.34, False
+    
+    # LondonStory model uses: LABEL_0 (Negative), LABEL_1 (Neutral), LABEL_2 (Positive)
     label_mapping = {
-        'negative': 'Negative',
-        'neutral': 'Neutral',
-        'positive': 'Positive',
         'LABEL_0': 'Negative',
         'LABEL_1': 'Neutral',
-        'LABEL_2': 'Positive'
+        'LABEL_2': 'Positive',
+        'negative': 'Negative',
+        'neutral': 'Neutral',
+        'positive': 'Positive'
     }
     
     for result in raw_results[0]:
-        label = result['label'].lower()
-        mapped_label = label_mapping.get(label, label_mapping.get(result['label'], 'Neutral'))
-        sentiment_scores[mapped_label] = result['score']
+        label = result['label']
+        score = result['score']
+        mapped_label = label_mapping.get(label, 'Neutral')
+        sentiment_scores[mapped_label] = score
     
     # Ensure all three sentiments exist
     for sentiment in ['Negative', 'Neutral', 'Positive']:
@@ -495,7 +501,7 @@ demo = gr.Interface(
     
     ### ✨ Advanced Features:
     - **🎙️ IndicWhisper ASR** - Specialized Hindi transcription model
-    - **🧠 XLM-RoBERTa** - Multilingual sentiment analysis
+    - **🧠 txlm-RoBERTa** - Hindi-optimized sentiment analysis
     - **🎵 Prosodic Analysis** - Voice tone, pitch, energy detection
     - **🔄 Mixed Emotion Detection** - Handles complex feelings
     - **🌐 Hinglish Support** - Works with Hindi + English mix