demo

app.py

@@ -4,14 +4,9 @@ import librosa
 import numpy as np
 import torch
 
-from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
-
-
-checkpoint = "microsoft/speecht5_tts"
-processor = SpeechT5Processor.from_pretrained(checkpoint)
-model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+from transformers import pipeline
 
+synthesiser = pipeline("text-to-speech", "techiaith/microsoft_speecht5_finetuned_bu_tts_cy_en")
 
 speaker_embeddings = {
     "GGP": "spkemb/speaker0.npy",
@@ -25,8 +20,6 @@ def predict(text, speaker):
     if len(text.strip()) == 0:
         return (16000, np.zeros(0).astype(np.int16))
 
-    inputs = processor(text=text, return_tensors="pt")
-
     # limit input length
     input_ids = inputs["input_ids"]
     input_ids = input_ids[..., :model.config.max_text_positions]
@@ -51,8 +44,9 @@ def predict(text, speaker):
 
     speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
 
-    speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
-
+    
+    speech = synthesiser(text, forward_params={"speaker_embeddings": speaker_embedding})
+    
     speech = (speech.numpy() * 32767).astype(np.int16)
     return (16000, speech)