Fix list to tensor performance and switch to soundfile to avoid torchcodec issue

app.py

@@ -18,19 +18,29 @@ def generate_amharic_speech(text: str, ref_audio_path: str = "ref_amharic.wav"):
         language="am"
     )
     
+    import numpy as np
+    import soundfile as sf
     output_path = "output.wav"
     
-    # Handle if generate() returns a list or tuple instead of a tensor
+    # Handle if generate() returns a tuple
     if isinstance(audio_tensor, tuple):
         audio_tensor = audio_tensor[0]
+        
+    # Convert to numpy array safely and handle lists of ndarrays
     if isinstance(audio_tensor, list):
-        audio_tensor = torch.tensor(audio_tensor)
+        audio_tensor = np.array(audio_tensor)
+        
+    if isinstance(audio_tensor, torch.Tensor):
+        audio_tensor = audio_tensor.cpu().numpy()
         
-    # torchaudio save expects a 2D tensor (channels, frames)
-    if audio_tensor.ndim == 1:
-        audio_tensor = audio_tensor.unsqueeze(0)
+    # soundfile expects shape (frames, channels) or (frames,)
+    # If the output is (channels, frames) like (1, N), we transpose it or squeeze it
+    if audio_tensor.ndim == 2 and audio_tensor.shape[0] < audio_tensor.shape[1]:
+        audio_tensor = audio_tensor.T
+    elif audio_tensor.ndim == 2 and audio_tensor.shape[0] == 1:
+        audio_tensor = audio_tensor.squeeze()
         
-    torchaudio.save(output_path, audio_tensor.cpu(), 24000)
+    sf.write(output_path, audio_tensor, 24000)
     return output_path
 
 # Expose Gradio Interface