app.py

import gradio as gr
import numpy as np
import torch
from datasets import load_dataset

from transformers import pipeline, VitsModel, VitsTokenizer

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)

def translate(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
    return outputs["text"]


# load translator to french
en_fr_translator = pipeline("translation_en_to_fr")


# load text-to-speech
model = VitsModel.from_pretrained("facebook/mms-tts-fra")
tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-fra")

def synthesise(text):
    translation_to_french = en_fr_translator(text)
    french_text = translation_to_french[0]['translation_text']

    inputs = tokenizer(french_text, return_tensors="pt")
    input_ids = inputs["input_ids"]
    
    with torch.no_grad():
        outputs = model(input_ids)
    
    speech = outputs["waveform"]
    return speech


# 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)

#embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
#speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)

#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()


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 French. 
Demo uses OpenAI's [Whisper Base](https://huggingface.co./openai/whisper-base) model for speech translation,
Google's [T5](https://huggingface.co./t5-base) for translating from English to French 
and Facebook's [Massive Multilingual Speech (MMS)](https://huggingface.co./facebook/mms-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")
"""

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="microphone", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    title=title,
    description=description,
)

file_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="upload", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    examples=[["./example.wav"]],
    title=title,
    description=description,
)

with demo:
    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])

demo.launch()