global pipeline

app.py

@@ -29,6 +29,16 @@ except subprocess.CalledProcessError:
     print("Warning: Could not install flash-attn, falling back to default attention")
 '''
 
+# Create global pipeline (similar to working HuggingFace example)
+pipe = pipeline(
+    "automatic-speech-recognition",
+    model="openai/whisper-large-v3-turbo",
+    torch_dtype=torch.float16,
+    device="cuda:0",
+    model_kwargs={"attn_implementation": "flash_attention_2"},
+    return_timestamps=True,
+)
+
 def comprehensive_flash_attention_verification():
     """Comprehensive verification of flash attention setup"""
     print("🔍 Running Flash Attention Verification...")
@@ -147,7 +157,7 @@ def comprehensive_flash_attention_verification():
 
 class WhisperTranscriber:
     def __init__(self):
-        self.pipe = None
+        self.pipe = pipe  # Use global pipeline
         self.diarization_model = None
         
     #@spaces.GPU
@@ -415,8 +425,8 @@ class WhisperTranscriber:
                 
             finally:
                 # Clean up temporary file
-                if os.path.exists(wav_path):
-                    os.unlink(wav_path)
+                if os.path.exists(audio_file):
+                    os.unlink(audio_file)
                     
         except Exception as e:
             import traceback
@@ -560,5 +570,4 @@ with demo:
     """)
 
 if __name__ == "__main__":
-    transcriber.setup_models()
     demo.launch(debug=True)