Update app.py

app.py

@@ -3,40 +3,46 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from transformers import AutoModel, AutoTokenizer, pipeline
+from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
-# Load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h", device=device)
+# 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")
 
-# Load text-to-speech checkpoint and speaker embeddings
-tokenizer = AutoTokenizer.from_pretrained("ttskit/ttskit-tts-ljspeech")
-model = AutoModel.from_pretrained("ttskit/ttskit-tts-ljspeech").to(device)
-vocoder = AutoModel.from_pretrained("ljspeech/vocoder-cryptron").to(device)
+model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
 
-# Sample code to load speaker embeddings (adjust according to the actual format of the dataset)
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
 speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
+
 def translate(audio):
     outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
     return outputs["text"]
 
+
 def synthesise(text):
-    inputs = tokenizer(text, return_tensors="pt")
+    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()
 
+
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
     synthesised_speech = synthesise(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
     return 16000, synthesised_speech
 
+
 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 Facebook's [Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base-960h) model for speech translation, ttskit's [ttskit-tts-ljspeech](https://huggingface.co/ttskit/ttskit-tts-ljspeech) for text-to-speech, and [Vocoder Cryptron](https://huggingface.co/ljspeech/vocoder-cryptron) for vocoding:
+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")
 """