app.py

import gradio as gr
import numpy as np
import torch
from datasets import load_dataset
from deep_translator import GoogleTranslator
from transformers import (
    AutoTokenizer,
    SpeechT5ForTextToSpeech,
    SpeechT5HifiGan,
    SpeechT5Processor,
    VitsModel,
    pipeline,
)

# device = "cuda:0" if torch.cuda.is_available() else "cpu"

device = "cpu"
# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition",
                    model="openai/whisper-base", device=device)

# load text-to-speech mms-tts-id model (speaker embeddings included)
model = VitsModel.from_pretrained("facebook/mms-tts-ind")
tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-ind")


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


def synthesise(text):
    inputs = tokenizer(text=text, return_tensors="pt")
    with torch.no_grad():
        speech = model(**inputs).waveform
    return speech.reshape(-1, 1).cpu()


def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    google_translated = GoogleTranslator(
        source="en", target="id").translate(translated_text)
    synthesised_speech = synthesise(google_translated)
    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 Indonesian. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech transcription, [Deep Translator](https://github.com/nidhaloff/deep-translator) for translation, and Meta's
[MMS TTS IND](https://huggingface.co/facebook/mms-tts-ind) 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()