Update app.py

app.py

@@ -26,13 +26,16 @@ from huggingface_hub import hf_hub_download
 
 from infer import load_model, eval_model
 from spkr import SpeakerEmbedding
-    
+
+spkr_model = None
+
 @spaces.GPU
 def spkr_model_init():
-    spkr_model = SpeakerEmbedding(device="cuda")
+    global spkr_model
+    if spkr_model is None:
+        spkr_model = SpeakerEmbedding(device="cuda")
     return spkr_model
 
-spkr_model = spkr_model_init()
 model, tokenizer, tokenizer_voila, model_type = load_model("maitrix-org/Voila-chat", "maitrix-org/Voila-Tokenizer")
 default_ref_file = "examples/character_ref_emb_demo.pkl"
 default_ref_name = "Homer Simpson"
@@ -49,6 +52,8 @@ default_ref_emb_mask_list = pickle.load(open(default_ref_file, "rb"))
 million_voice_ref_emb_mask_list = pickle.load(open(million_voice_ref_file, "rb"))
 
 def get_ref_embs(ref_audio):
+    global spkr_model
+    spkr_model = spkr_model_init()
     wav, sr = torchaudio.load(ref_audio)
     ref_embs = spkr_model(wav, sr).cpu()
     return ref_embs