Update fish_speech/inference_engine/reference_loader.py

fish_speech/inference_engine/reference_loader.py

@@ -1,3 +1,4 @@
+
 import io
 from hashlib import sha256
 from pathlib import Path
@@ -32,11 +33,15 @@ class ReferenceLoader:
         self.encode_reference: Callable
 
         # Define the torchaudio backend
-        backends = torchaudio.list_audio_backends()
-        if "ffmpeg" in backends:
-            self.backend = "ffmpeg"
+        if hasattr(torchaudio, "list_audio_backends"):
+            backends = torchaudio.list_audio_backends()
+            if "ffmpeg" in backends:
+                self.backend = "ffmpeg"
+            else:
+                self.backend = "soundfile"
         else:
-            self.backend = "soundfile"
+            # For newer versions of torchaudio, let it auto-detect the backend
+            self.backend = None
 
     def load_by_id(
         self,
@@ -130,3 +135,4 @@ class ReferenceLoader:
 
         audio = waveform.squeeze().numpy()
         return audio
+