Update app.py

app.py

@@ -2,6 +2,7 @@ import gradio as gr
 import numpy as np
 import torch
 from datasets import load_dataset
+from transformers import VitsModel, VitsTokenizer
 
 from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
 
@@ -10,29 +11,23 @@ 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)
+translator = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ru")
 
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("sanchit-gandhi/speecht5_tts_vox_nl").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+model = VitsModel.from_pretrained('facebook/mms-tts-rus').to(device)
+tokenizer = VitsTokenizer.from_pretrained('facebook/mms-tts-rus')
 
-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", "language": "dutch"})
-    return outputs["text"]
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    return translator(outputs['text'])[0]['translation_text']
 
 
 def synthesise(text):
-    inputs = processor(text=text, return_tensors="pt",
-                       max_length=582,
-                       truncation=True,
-                      )
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    inputs = tokenizer(text=text, return_tensors="pt")
+    with torch.no_grad():
+        speech = model(**inputs).waveform
     return speech.cpu()
 
 
@@ -40,13 +35,12 @@ 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
+    return 16000, synthesised_speech[0]
 
 
 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:
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """
 
@@ -68,8 +62,3 @@ file_translate = gr.Interface(
     title=title,
     description=description,
 )
-
-with demo:
-    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
-
-demo.launch(share=True)