Update app.py

app.py

@@ -4,9 +4,24 @@ from transformers import pipeline
 from jiwer import wer
 
 # Load models
-whisper_pipeline_1 = pipeline("automatic-speech-recognition", model="maliahson/Finetuned_Whisper_Medium_Model_2")
-whisper_pipeline_2 = pipeline("automatic-speech-recognition", model="openai/whisper-large-v3-turbo", device=0 if torch.cuda.is_available() else "cpu")
-whisper_pipeline_3 = pipeline("automatic-speech-recognition", model="openai/whisper-medium", device=0 if torch.cuda.is_available() else "cpu")
+whisper_pipeline_1 = pipeline(
+    "automatic-speech-recognition", 
+    model="maliahson/Finetuned_Whisper_Medium_Model_2"
+)
+
+whisper_pipeline_2 = pipeline(
+    "automatic-speech-recognition", 
+    model="openai/whisper-large-v3-turbo", 
+    device=0 if torch.cuda.is_available() else "cpu"
+)
+
+whisper_pipeline_3 = pipeline(
+    "automatic-speech-recognition", 
+    model="openai/whisper-medium", 
+    device=0 if torch.cuda.is_available() else "cpu", 
+    # Explicitly set language and task for Urdu transcription
+    model_kwargs={"language": "<|ur|>", "task": "transcribe"}
+)
 
 def transcribe_and_compare(audio_path, original_transcription=None):
     """
@@ -18,27 +33,30 @@ def transcribe_and_compare(audio_path, original_transcription=None):
     Returns:
         dict: Results including transcriptions and WER calculations.
     """
+    # Transcriptions from all three models
     transcription_1 = whisper_pipeline_1(audio_path)["text"]
     transcription_2 = whisper_pipeline_2(audio_path)["text"]
     transcription_3 = whisper_pipeline_3(audio_path)["text"]
     
+    # Prepare comparison results
     comparison_result = {
         "Model 1 Output (maliahson/Finetuned_Whisper_Medium_Model_2)": transcription_1,
         "Model 2 Output (openai/whisper-large-v3-turbo)": transcription_2,
-        "Model 3 Output (Openai/whisper-medium)": transcription_3
+        "Model 3 Output (Openai/whisper-medium, Urdu)": transcription_3
     }
 
     if original_transcription:
-        # Calculate Word Error Rate (WER)
+        # Calculate Word Error Rate (WER) for all models
         wer_1 = wer(original_transcription, transcription_1)
         wer_2 = wer(original_transcription, transcription_2)
         wer_3 = wer(original_transcription, transcription_3)
         
+        # Add WER scores to results
         comparison_result["WER Model 1"] = wer_1
         comparison_result["WER Model 2"] = wer_2
         comparison_result["WER Model 3"] = wer_3
     else:
-        # Compare outputs of all three models
+        # Compare outputs of all three models when no ground truth is provided
         comparison_result["Difference Between Models"] = {
             "Model 1 Unique Words": set(transcription_1.split()) - set(transcription_2.split()) - set(transcription_3.split()),
             "Model 2 Unique Words": set(transcription_2.split()) - set(transcription_1.split()) - set(transcription_3.split()),