rev app

app.py

@@ -1,49 +1,36 @@
+import spaces
 import gradio as gr
 from transformers import pipeline
-import torch
-import spaces
 
-# Initialize model on CPU
+# Initialize model and move to GPU
 model = pipeline(
     "automatic-speech-recognition",
     model="Aekanun/whisper-small-hi",
-    device="cpu"
+    device="cuda"  # เปลี่ยนเป็น cuda เลย
 )
 
-@spaces.GPU
+@spaces.GPU  # GPU function with default 60s duration
 def transcribe_speech(audio):
-    """Speech transcription with GPU support"""
+    """Speech transcription function"""
     try:
         if audio is None:
             return "กรุณาบันทึกเสียงก่อน"
             
-        # Move model to GPU
-        model.model = model.model.to("cuda")
+        # Process audio (model is already on GPU)
+        result = model(audio, batch_size=1)
         
-        # Make sure input is on the same device as model
-        with torch.cuda.amp.autocast():
-            # Process audio
-            result = model(audio, batch_size=1)
-            
-            # Get text result
-            text = result["text"] if isinstance(result, dict) else result
-            
-        # Move model back to CPU
-        model.model = model.model.to("cpu")
-        torch.cuda.empty_cache()
+        # Get text result
+        text = result["text"] if isinstance(result, dict) else result
         
         return text
         
     except Exception as e:
-        # Make sure model is back on CPU in case of error
-        model.model = model.model.to("cpu")
-        torch.cuda.empty_cache()
         return f"เกิดข้อผิดพลาด: {str(e)}"
 
 # Create Gradio interface
 demo = gr.Interface(
     fn=transcribe_speech,
-    inputs=gr.Audio(type="filepath"),  # Simplified Audio component
+    inputs=gr.Audio(type="filepath"),
     outputs=gr.Textbox(label="ข้อความ"),
     title="Thai Speech Transcription",
     description="บันทึกเสียงเพื่อแปลงเป็นข้อความภาษาไทย",