Update app.py

app.py

@@ -2,18 +2,19 @@ import gradio as gr
 from faster_whisper import WhisperModel
 
 model_size = 'large-v3'
-model = WhisperModel(model_size, device="auto", compute_type="float16")
 
-# or run on GPU with INT8
-# model = WhisperModel(model_size, device="cuda", compute_type="int8_float16")
-# or run on CPU with INT8
-# model = WhisperModel(model_size, device="cpu", compute_type="int8")
+def load_model(model_size):
+    if torch.cuda.is_available():
+        model = WhisperModel(model_size, device="cuda", compute_type="float16")
+        # model = WhisperModel(model_size, device="cuda", compute_type="int8_float16")
+    else:
+        model = WhisperModel(model_size, device="cpu", compute_type="int8")
 
 def speech_to_text(audio_file, _model_size):
     global model_size, model
     if model_size != _model_size:
         model_size = _model_size
-        model = WhisperModel(model_size, device="auto", compute_type="float16")
+        model = load_model(model_size)
 
     with torch.no_grad():
         segments, info = model.transcribe(
@@ -29,10 +30,13 @@ def speech_to_text(audio_file, _model_size):
     for segment in segments:
         text += "{segment.start:.2f}\t{segment.end:.2f}\t{segment.text}\n"
 
+
+load_model(model_size)
+
 gr.Interface(
     fn=speech_to_text,
     inputs=[
-        gr.Audio(source="upload", type="filepath"),
+        gr.Audio(sources="upload", type="filepath"),
         gr.Dropdown(value=model_size, choices=["tiny", "base", "small", "medium", "large", "large-v2", "large-v3"]),
         ],
     outputs="text").launch()