app.py

import gradio as gr
import numpy as np
import torch

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 translation checkpoint
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-tc-big-en-pt")

#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):
    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 = 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):
    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 - English to Portuguese"
description = """
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")
"""

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(sources="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(sources="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()