ugh

vall_e/config.py

@@ -839,6 +839,8 @@ class Inference:
 
 	batch_size: int = 16 # I don't know what would be a good batch size
 
+	audio_backends: dict = field(default_factory=lambda: {})
+
 	@property
 	def dtype(self):
 		return coerce_dtype(self.weight_dtype)

vall_e/emb/qnt.py

@@ -22,30 +22,30 @@ from torch.nn.utils.rnn import pad_sequence
 
 try:
 	from .codecs.encodec import *
+	cfg.inference.audio_backends["encodec"] = True
 except Exception as e:
-	cfg.inference.use_encodec = False
+	cfg.inference.audio_backends["encodec"] = False # e
-	#raise e
 	_logger.warning(str(e))
 
 try:
 	from .codecs.vocos import *
+	cfg.inference.audio_backends["vocos"] = True
 except Exception as e:
-	cfg.inference.use_vocos = False
+	cfg.inference.audio_backends["vocos"] = False # e
-	#raise e
 	_logger.warning(str(e))
 
 try:
 	from .codecs.dac import *
+	cfg.inference.audio_backends["dac"] = True
 except Exception as e:
-	cfg.inference.use_dac = False
+	cfg.inference.audio_backends["dac"] = False # e
-	#raise e
 	_logger.warning(str(e))
 
 try:
 	from .codecs.nemo import *
+	cfg.inference.audio_backends["nemo"] = True
 except Exception as e:
-	cfg.inference.use_nemo = False
+	cfg.inference.audio_backends["nemo"] = False # e
-	#raise e
 	_logger.warning(str(e))
 
 @cache

vall_e/inference.py

@@ -62,7 +62,7 @@ class TTS():
 		cfg.dataset.use_hdf5 = False # could use cfg.load_hdf5(), but why would it ever need to be loaded for inferencing
 
 		# fallback to encodec if no vocos
-		if cfg.audio_backend == "vocos" and not cfg.inference.use_vocos:
+		if cfg.audio_backend == "vocos" and not cfg.inference.audio_backends.get("vocos", False):
 			_logger.warning("Vocos requested but not available, falling back to Encodec...")
 			cfg.set_audio_backend(cfg.audio_backend)