use soudfile for speaker sep audio saving

gradio_app.py

@@ -3,6 +3,7 @@ import uuid
 import torch
 import torchaudio
 import torchaudio.transforms as T
+import soundfile as sf
 import gradio as gr
 import look2hear.models
 
@@ -65,7 +66,8 @@ def separate_speakers(audio_path):
     output_files = []
     for i in range(ests_speech.shape[0]):
         path = os.path.join(output_dir, f"speaker_{i+1}.wav")
-        torchaudio.save(path, ests_speech[i].cpu(), TARGET_SR)
+        audio_np = ests_speech[i].cpu().numpy()
+        sf.write(path, audio_np.T, TARGET_SR)  # Transpose only if shape is [T, C], usually not needed
         output_files.append(path)
 
     updates = []