diff --git a/requirements.txt b/requirements.txt
index 630820c..1d8c3b3 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -3,3 +3,4 @@ soundfile
 python-vlc
 tqdm
 argparse
+torch
diff --git a/tts.py b/tts.py
index aeac894..dc7bffa 100644
--- a/tts.py
+++ b/tts.py
@@ -1,7 +1,8 @@
 import sys
 import os
-from time import sleep
+from time import sleep, time
 
+import torch
 import argparse
 from kokoro import KPipeline
 import soundfile as sf
@@ -32,11 +33,17 @@ def parse_args():
         default="demo/tongue-twister.txt",
         help="Voice to use (pro, hot, asmr, brit)",
     )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default=("cuda" if torch.cuda.is_available() else ("mps" if torch.backends.mps.is_available() else ("xpu" if torch.xpu.is_available() else "cpu"))),
+        help="Device for inference: cuda | mps | cpu",
+    )
     return parser.parse_args()
 
 def main():
     args=parse_args()
-    pipeline = KPipeline(lang_code='a', device='xpu', repo_id='hexgrad/Kokoro-82M')
+    pipeline = KPipeline(lang_code='a', device=args.device, repo_id='hexgrad/Kokoro-82M')
     voice=voices[args.voice]
     if voice is None:
         if args.voice is None:
@@ -57,6 +64,7 @@ def main():
     output_files = []
     length = 0
 
+    start_time = time()
     for i, (gs, ps, audio) in enumerate(generator):
         output_file_name=f'outputs/{name}-{i}.wav'
         os.makedirs(os.path.dirname(output_file_name), exist_ok=True)
@@ -64,6 +72,8 @@ def main():
         sf.write(output_file_name, audio, 24000)
         print(u'\u2713', output_file_name)
         length = length + 1
+    generation_time = time() - start_time
+    print(f"Generation time: {generation_time:.2f} seconds")
 
     for i, output in enumerate(output_files):
         full_path = os.path.abspath(output)