Update app.py

app.py

@@ -151,8 +151,12 @@ def generate_speaker_clips(annotations, waveform, sample_rate, file_index):
 
         clip_waveform = waveform[:, start_sample:end_sample]
 
+        # Use soundfile (pure C, no FFmpeg/torchcodec) to encode WAV bytes
+        # soundfile expects (samples, channels) numpy array; waveform is (channels, samples) tensor
+        clip_np = clip_waveform.numpy().T
         buffer = io.BytesIO()
-        torchaudio.save(buffer, clip_waveform, sample_rate, format="wav")
+        import soundfile as sf
+        sf.write(buffer, clip_np, sample_rate, format="WAV", subtype="PCM_16")
         buffer.seek(0)
         clips[speaker] = buffer.read()