Update app.py

app.py

@@ -3,8 +3,15 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
-
+from transformers import (
+    SpeechT5ForTextToSpeech,
+    SpeechT5HifiGan,
+    SpeechT5Processor,
+    VitsModel,
+    VitsTokenizer,
+    pipeline
+)
+from transformers import VitsModel, VitsTokenizer
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
@@ -14,7 +21,10 @@ asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base",
 # load text-to-speech checkpoint and speaker embeddings
 processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
+# model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
+model = VitsModel.from_pretrained("Matthijs/mms-tts-nld")
+tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-nld")
+
 vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
@@ -27,9 +37,18 @@ def translate(audio):
 
 
 def synthesise(text):
-    inputs = processor(text=text, return_tensors="pt")
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
-    return speech.cpu()
+    # inputs = processor(text=text, return_tensors="pt")
+    # speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    # speech = speech.cpu()
+    
+    inputs = tokenizer(text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
+
+    with torch.no_grad():
+        outputs = model(input_ids)
+
+    speech = outputs.audio[0].cpu()
+    return speech
 
 
 def speech_to_speech_translation(audio):
@@ -43,7 +62,6 @@ title = "Cascaded STST"
 description = """
 Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
 [SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
-
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """
 
@@ -69,4 +87,4 @@ file_translate = gr.Interface(
 with demo:
     gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
 
-demo.launch()
+demo.launch()