Skip to content

SOTA Open Source TTS

fix: add torchaudio 2.9 compatibility for list_audio_backends removal (#1148)

Harikrishna KP75a9afe46ac6
diff --git a/fish_speech/inference_engine/reference_loader.py b/fish_speech/inference_engine/reference_loader.py
index 8fa6817..96a8e8b 100644
--- a/fish_speech/inference_engine/reference_loader.py
+++ b/fish_speech/inference_engine/reference_loader.py
@@ -31,11 +31,22 @@ 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"
+ # list_audio_backends() was removed in torchaudio 2.9
+ try:
+ backends = torchaudio.list_audio_backends()
+ if "ffmpeg" in backends:
+ self.backend = "ffmpeg"
+ else:
+ self.backend = "soundfile"
+ except AttributeError:
+ # torchaudio 2.9+ removed list_audio_backends()
+ # Try ffmpeg first, fallback to soundfile
+ try:
+ import torchaudio.io._load_audio_fileobj # noqa: F401
+
+ self.backend = "ffmpeg"
+ except (ImportError, ModuleNotFoundError):
+ self.backend = "soundfile"
def load_by_id(
self,
diff --git a/tools/vqgan/extract_vq.py b/tools/vqgan/extract_vq.py
index d50748c..e1c743e 100644
--- a/tools/vqgan/extract_vq.py
+++ b/tools/vqgan/extract_vq.py
@@ -23,12 +23,22 @@ OmegaConf.register_new_resolver("eval", eval)
# This file is used to convert the audio files to text files using the Whisper model.
# It's mainly used to generate the training data for the VQ model.
-backends = torchaudio.list_audio_backends()
-
-if "ffmpeg" in backends:
- backend = "ffmpeg"
-else:
- backend = "soundfile"
+# Determine audio backend - list_audio_backends() was removed in torchaudio 2.9
+try:
+ backends = torchaudio.list_audio_backends()
+ if "ffmpeg" in backends:
+ backend = "ffmpeg"
+ else:
+ backend = "soundfile"
+except AttributeError:
+ # torchaudio 2.9+ removed list_audio_backends()
+ # Try ffmpeg first, fallback to soundfile
+ try:
+ import torchaudio.io._load_audio_fileobj # Check if ffmpeg backend is available
+
+ backend = "ffmpeg"
+ except (ImportError, ModuleNotFoundError):
+ backend = "soundfile"
RANK = int(os.environ.get("SLURM_PROCID", 0))
WORLD_SIZE = int(os.environ.get("SLURM_NTASKS", 1))