Update app.py

app.py

@@ -1,35 +1,40 @@
-import gradio as gr
+mport gradio as gr
 import numpy as np
 import torch
-from datasets import load_dataset
+import scipy.io.wavfile
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from datasets import load_dataset
 
+from transformers import pipeline, VitsModel, AutoTokenizer
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 # load speech translation checkpoint
 asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
 
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
-
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+#load translation checkpoint
+translator = pipeline("translation", model="Helsinki-NLP/opus-mt-tc-big-en-pt")
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation", trust_remote_code=True)
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+#load tts model to portuguese and tokenizer
+tts_model_name = "facebook/mms-tts-por"
+tts_model = VitsModel.from_pretrained(tts_model_name)
+tts_tokenizer = AutoTokenizer.from_pretrained(tts_model_name)
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, generate_kwargs={"task": "translate"})
-    return outputs["text"]
+    transcribed_outputs = asr_pipe(audio, generate_kwargs={"task": "translate"})
+    transcribed_text = transcribed_outputs["text"]
+    outputs = translator(transcribed_text)
+    return outputs[0]["translation_text"]
 
 
 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 = tts_tokenizer(text=text, return_tensors="pt")
+    input_ids = inputs["input_ids"].to(device)
+    with torch.no_grad():
+        speech = tts_model(input_ids).waveform
+    speech = speech.cpu().squeeze()  # Remove dimensões extras, resultando em tensor 1D
+    return speech
 
 
 def speech_to_speech_translation(audio):
@@ -39,10 +44,10 @@ def speech_to_speech_translation(audio):
     return 16000, synthesised_speech
 
 
-title = "Cascaded STST"
+title = "Cascaded STST - English to Portuguese"
 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:
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Portuguese. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Meta's
+[MMS-TTS-POR](https://huggingface.co/facebook/mms-tts-por) 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")
 """