app.py

import gradio as gr
from transformers import pipeline
import torch
import numpy as np

device = "cuda:0" if torch.cuda.is_available() else "cpu"
device = 'cpu'
print("Device:", device)

pipe_translate = pipeline("translation", model="Helsinki-NLP/opus-mt-fr-en", device=device)
pipe_tts = pipeline("text-to-speech", model="facebook/mms-tts-eng", device=device) # Better quality, way faster than bark

def get_translation(text):
    return pipe_translate(text)[0]["translation_text"]

def get_audio(text):
    speech = pipe_tts(text)
    return speech["sampling_rate"], (speech["audio"]* 32767).astype(np.int16).T

with gr.Blocks() as demo:    
    input_text = gr.Textbox(
                label="Input text",
                info="Your text",
                lines=3,
                placeholder="Écrire le texte à traduire",
            )

    
    
    translation_button = gr.Button("Traduire...")
    output_text = gr.Textbox(
                label="Output text",
                info="Your text",
                lines=3,
                placeholder="Votre traduction",
            )
    speech_button = gr.Button("Générer audio...")
    translation_button.click(
            get_translation,
            inputs=[
                input_text
            ],
            outputs=[
                output_text
            ],
    )
    speech_button.click(
            get_audio,
            inputs=[
                output_text
            ],
            outputs=[
                gr.Audio(label="Output")
            ],
    )

demo.launch()