diff --git a/transcribe_all.py b/transcribe_all.py
index de02fa6..e8df9af 100755
--- a/transcribe_all.py
+++ b/transcribe_all.py
@@ -1,6 +1,7 @@
 import os
 import sys
 import time
+import torch
 import whisper
 import concurrent.futures
 import json
@@ -18,7 +19,6 @@ with open("transcription_config.yml", "r", encoding="utf-8") as file:
     settings = yaml.safe_load(file)
     folder_list = settings.get("folder_list")
     model_name = settings.get("model_name")
-    device = settings.get("device")
 
 def load_audio_librosa(path: str, sr: int = 16_000) -> np.ndarray:
     audio, orig_sr = librosa.load(path, sr=sr)      # load + resample to 16 kHz
@@ -222,7 +222,10 @@ def process_folder(root_folder):
     else:
         print(f"Checked {checked_files} files. Start to transcribe {len(valid_files)} files.")
     
-    print("Loading Whisper model...")
+    # Choose “cuda” if available, otherwise “cpu”
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"Loading Whisper model on {device}…")
+
     model = whisper.load_model(model_name, device=device)
     
     # Use a thread pool to pre-load files concurrently.