demo

app.py

@@ -4,10 +4,11 @@ import librosa
 import numpy as np
 import torch
 
-from transformers import pipeline
+from transformers import SpeechT5ForTextToSpeech, SpeechT5Processor, SpeechT5HifiGan
 
-synthesiser = pipeline("text-to-speech", "techiaith/microsoft_speecht5_finetuned_bu_tts_cy_en", from_tf=True)
-synthesiser.to('cuda')
+checkpoint = "microsoft/speecht5_tts"
+processor = SpeechT5Processor.from_pretrained(checkpoint)
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
 
 speaker_embeddings = {
     "GGP": "spkemb/speaker0.npy",
@@ -15,7 +16,6 @@ speaker_embeddings = {
     "BDP": "spkemb/speaker2.npy",
 }
 
-
 @spaces.GPU
 def predict(text, speaker):
     if len(text.strip()) == 0:
@@ -23,7 +23,12 @@ def predict(text, speaker):
 
     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})
+    model = SpeechT5ForTextToSpeech.from_pretrained(
+        "techiaith/microsoft_speecht5_finetuned_bu_tts_cy_en"
+    )
+
+    inputs = processor(text=text, return_tensors="pt")
+    speech = model.generate_speech(inputs["input_ids"], speaker_embedding, vocoder=vocoder)
     speech = (speech.numpy() * 32767).astype(np.int16)
     return (16000, speech)