Update app.py

app.py

@@ -49,18 +49,20 @@ def load_models():
     # Load IndicWhisper for Hindi ASR
     print("🎤 Loading IndicWhisper Hindi ASR model...")
     try:
-        ASR_PROCESSOR = AutoProcessor.from_pretrained("vasista22/whisper-hindi-medium")
-        ASR_MODEL = AutoModelForSpeechSeq2Seq.from_pretrained("vasista22/whisper-hindi-medium")
-        
-        # Create pipeline with the loaded model
+        device = "cuda:0" if torch.cuda.is_available() else "cpu"
         ASR_PIPELINE = pipeline(
             "automatic-speech-recognition",
-            model=ASR_MODEL,
-            tokenizer=ASR_PROCESSOR.tokenizer,
-            feature_extractor=ASR_PROCESSOR.feature_extractor,
-            device="cpu",
-            chunk_length_s=30
+            model="vasista22/whisper-hindi-medium",
+            chunk_length_s=30,
+            device=device
+        )
+        
+        # FIX: Set forced_decoder_ids properly for the model config
+        ASR_PIPELINE.model.config.forced_decoder_ids = ASR_PIPELINE.tokenizer.get_decoder_prompt_ids(
+            language="hi", 
+            task="transcribe"
         )
+        
         print("✅ IndicWhisper Hindi ASR model loaded successfully")
     except Exception as e:
         print(f"❌ Error loading IndicWhisper, trying fallback: {e}")
@@ -360,8 +362,7 @@ def predict(audio_filepath):
         # Validation
         if audio_filepath is None:
             return {
-                "⚠️ Error": 1.0,
-                "Message": "No audio file uploaded"
+                "⚠️ Error": "No audio file uploaded"
             }
         
         # ============================================
@@ -384,13 +385,8 @@ def predict(audio_filepath):
         # ============================================
         print("🔄 Transcribing with cached IndicWhisper model...")
         try:
-            result = ASR_PIPELINE(
-                audio_filepath,
-                generate_kwargs={
-                    "language": "hindi",
-                    "task": "transcribe"
-                }
-            )
+            # FIX: Don't pass language in generate_kwargs, it's already set in model config
+            result = ASR_PIPELINE(audio_filepath)
             
             transcription = result["text"].strip()
             print(f"📝 Transcription: '{transcription}'")
@@ -398,8 +394,7 @@ def predict(audio_filepath):
         except Exception as asr_error:
             print(f"❌ ASR Error: {asr_error}")
             return {
-                "⚠️ ASR Error": 1.0,
-                "Message": str(asr_error)
+                "⚠️ ASR Error": str(asr_error)
             }
         
         # ============================================
@@ -407,8 +402,7 @@ def predict(audio_filepath):
         # ============================================
         if not transcription or len(transcription) < 2:
             return {
-                "⚠️ No Speech Detected": 1.0,
-                "Transcription": transcription or "Empty"
+                "⚠️ No Speech Detected": f"Transcription: {transcription or 'Empty'}"
             }
         
         is_valid, validation_msg, hindi_ratio = validate_hindi_text(transcription)
@@ -416,9 +410,8 @@ def predict(audio_filepath):
         
         if not is_valid:
             return {
-                "⚠️ Language Error": 1.0,
-                "Message": validation_msg,
-                "Transcription": transcription
+                "⚠️ Language Error": validation_msg,
+                "📝 Transcription": transcription
             }
         
         # ============================================
@@ -435,19 +428,25 @@ def predict(audio_filepath):
             )
             
             # ============================================
-            # STEP 5: Format Results
+            # STEP 5: Format Results (FIX: All values must be float)
             # ============================================
             result_dict = {}
             
+            # Add sentiment scores (all floats)
             for sentiment, score in sorted(sentiment_scores.items(), key=lambda x: x[1], reverse=True):
                 result_dict[f"{sentiment}"] = float(score)
             
-            result_dict["📝 Transcription"] = transcription
-            result_dict["🎯 Confidence"] = float(confidence)
-            result_dict["🔀 Mixed Emotions"] = "Yes" if is_mixed else "No"
-            result_dict["🌐 Hindi Content"] = f"{hindi_ratio*100:.0f}%"
+            # FIX: Convert all metadata to float values for compatibility
+            # Use very small values to put them at the bottom of the sorted list
+            result_dict["_Confidence"] = float(confidence)
+            result_dict["_Mixed_Emotions"] = 1.0 if is_mixed else 0.0
+            result_dict["_Hindi_Content_Pct"] = float(hindi_ratio * 100)
             
+            # Store transcription separately for display
+            print(f"📝 Full Transcription: {transcription}")
             print(f"✅ Complete! Confidence: {confidence:.3f}")
+            print(f"🔀 Mixed Emotions: {'Yes' if is_mixed else 'No'}")
+            print(f"🌐 Hindi Content: {hindi_ratio*100:.0f}%")
             print(f"{'='*60}\n")
             
             return result_dict
@@ -455,9 +454,7 @@ def predict(audio_filepath):
         except Exception as sentiment_error:
             print(f"❌ Sentiment Error: {sentiment_error}")
             return {
-                "⚠️ Sentiment Error": 1.0,
-                "Message": str(sentiment_error),
-                "Transcription": transcription
+                "⚠️ Sentiment Error": str(sentiment_error)
             }
         
     except Exception as e:
@@ -465,8 +462,7 @@ def predict(audio_filepath):
         import traceback
         traceback.print_exc()
         return {
-            "⚠️ System Error": 1.0,
-            "Message": str(e)
+            "⚠️ System Error": str(e)
         }
 
 # ============================================
@@ -508,10 +504,10 @@ demo = gr.Interface(
     
     ### 📊 Output Includes:
     - Sentiment probabilities (Positive/Negative/Neutral)
-    - Exact transcription in Hindi/Devanagari
-    - Confidence score (how sure the model is)
-    - Mixed emotion indicator
-    - Language composition (% Hindi content)
+    - _Confidence: Prediction confidence score
+    - _Mixed_Emotions: 1.0 if mixed, 0.0 if not
+    - _Hindi_Content_Pct: Percentage of Hindi characters
+    - Check console logs for full transcription
     
     ### 💡 Best Practices:
     1. Speak clearly for 3-10 seconds