demo

app.py

@@ -21,33 +21,9 @@ def predict(text, speaker):
     if len(text.strip()) == 0:
         return (16000, np.zeros(0).astype(np.int16))
 
-    # limit input length
-    input_ids = inputs["input_ids"]
-    input_ids = input_ids[..., :model.config.max_text_positions]
-
-    if speaker == "Surprise Me!":
-        # load one of the provided speaker embeddings at random
-        idx = np.random.randint(len(speaker_embeddings))
-        key = list(speaker_embeddings.keys())[idx]
-        speaker_embedding = np.load(speaker_embeddings[key])
-
-        # randomly shuffle the elements
-        np.random.shuffle(speaker_embedding)
-
-        # randomly flip half the values
-        x = (np.random.rand(512) >= 0.5) * 1.0
-        x[x == 0] = -1.0
-        speaker_embedding *= x
-
-        #speaker_embedding = np.random.rand(512).astype(np.float32) * 0.3 - 0.15
-    else:
-        speaker_embedding = np.load(speaker_embeddings[speaker[:3]])
-
+    speaker_embedding = np.load(speaker_embeddings[speaker[:3]])
     speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
-
-    
     speech = synthesiser(text, forward_params={"speaker_embeddings": speaker_embedding})
-    
     speech = (speech.numpy() * 32767).astype(np.int16)
     return (16000, speech)
 
@@ -72,7 +48,6 @@ gr.Interface(
             "GGP (gwryw-gogledd-pro)",
             "BGP (benyw-gogledd-pro)",
             "BDP (benyw-de-pro)",
-            "Surprise Me!"
         ],
         value="GGP (gwryw-gogledd-pro)"),
     ],