rev app

app.py

@@ -1,30 +1,47 @@
-import spaces
 import gradio as gr
 from transformers import pipeline
+import torch
+import spaces
 
-# Initialize model and move to GPU
+# Initialize model on CPU
 model = pipeline(
     "automatic-speech-recognition",
     model="Aekanun/whisper-small-hi",
-    device="cuda"  # เปลี่ยนเป็น cuda เลย
+    device="cpu"
 )
 
-@spaces.GPU  # GPU function with default 60s duration
+@spaces.GPU
 def transcribe_speech(audio):
-    """Speech transcription function"""
+    """Speech transcription with GPU support"""
     try:
         if audio is None:
             return "กรุณาบันทึกเสียงก่อน"
             
-        # Process audio (model is already on GPU)
-        result = model(audio, batch_size=1)
+        # Move model to GPU
+        model.model = model.model.to("cuda")
         
-        # Get text result
-        text = result["text"] if isinstance(result, dict) else result
+        with torch.amp.autocast('cuda'):
+            # Process audio with chunk length to handle long audio
+            result = model(
+                audio,
+                batch_size=1,
+                return_timestamps=True,  # แก้ error เรื่อง timestamps
+                chunk_length_s=30        # แก้ error เรื่องความยาวเสียง
+            )
+            
+            # 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()
         
         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