Update app.py

app.py

@@ -1,10 +1,4 @@
-import gradio as gr
-from transformers import pipeline
-import torch
 import torchaudio
-import numpy as np
-
-device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 def transcribe(audio):
     # Extract the sample rate and audio data from the tuple
@@ -14,13 +8,20 @@ def transcribe(audio):
     if not isinstance(audio_data, np.ndarray):
         audio_data = np.array(audio_data)
     
+    # Convert to a tensor
+    audio_tensor = torch.tensor(audio_data)
+    
     # Convert to mono if the audio is stereo
-    if audio_data.ndim > 1:
-        audio_data = np.mean(audio_data, axis=0)
+    if audio_tensor.ndim > 1:
+        audio_tensor = torch.mean(audio_tensor, dim=0, keepdim=True)
+    
+    # Resample to 16kHz if necessary
+    if sample_rate != 16000:
+        resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
+        audio_tensor = resampler(audio_tensor)
     
-    # Reshape the audio data to match the expected input format (1, num_samples)
-    if audio_data.ndim == 1:
-        audio_data = np.expand_dims(audio_data, axis=0)
+    # Ensure the audio tensor is on the correct device
+    audio_tensor = audio_tensor.to(device)
     
     pipe = pipeline(
         "automatic-speech-recognition",
@@ -29,15 +30,5 @@ def transcribe(audio):
         device=device,
     )
     
-    prediction = pipe(audio_data)["text"]
+    prediction = pipe(audio_tensor)["text"]
     return prediction
-
-gradio_app = gr.Interface(
-    fn=transcribe,
-    inputs=gr.Audio(label="Input"),
-    outputs=gr.Textbox(label="Result"),
-    title="Transcribed",
-)
-
-if __name__ == "__main__":
-    gradio_app.launch()