torchaudio 2.11 removed list_audio_backends() along with the whole
sox/ffmpeg backend layer; its load() delegates to torchcodec and
accepts-but-ignores backend=. Select statically what 2.11 ignores.
--- a/fish_speech/inference_engine/reference_loader.py
+++ b/fish_speech/inference_engine/reference_loader.py
@@ -31,11 +31,9 @@ class ReferenceLoader:
         self.encode_reference: Callable

         # Define the torchaudio backend
-        backends = torchaudio.list_audio_backends()
-        if "ffmpeg" in backends:
-            self.backend = "ffmpeg"
-        else:
-            self.backend = "soundfile"
+        # Gentoo: torchaudio 2.11 removed list_audio_backends(); its
+        # load() delegates to torchcodec and ignores backend=.
+        self.backend = "soundfile"

     def load_by_id(
         self,