app.py

from vietTTS.hifigan.mel2wave import mel2wave
from vietTTS.nat.text2mel import text2mel
from vietTTS import nat_normalize_text
import numpy as np
import gradio as gr
import os


def download_assets():
    os.system("gdown --id 16UhN8QBxG1YYwUh8smdEeVnKo9qZhvZj -O duration_latest_ckpt.pickle")
    os.system("gdown --id 1-8Ig65S3irNHSzcskT37SLgeyuUhjKdj -O acoustic_latest_ckpt.pickle")
    os.system("gdown --id 19cRNDC6IrHFAAE4U9I7K0mzLMgPsi5zb -O hk_hifi.pickle")
    os.system("wget https://raw.githubusercontent.com/NTT123/vietTTS/master/assets/hifigan/config.json")
    os.system("wget https://raw.githubusercontent.com/NTT123/vietTTS/master/assets/infore/lexicon.txt")

def text_to_speech(text):
    # prevent too long text
    if len(text) > 500:
        text = text[:500]
    text = nat_normalize_text(text)
    mel = text2mel(
        text,
        "lexicon.txt",
        0.2,
        "acoustic_latest_ckpt.pickle",
        "duration_latest_ckpt.pickle",
    )
    wave = mel2wave(mel, "config.json", "hk_hifi.pickle")
    return (wave * (2**15)).astype(np.int16)


def speak(text):
    y = text_to_speech(text)
    return 16_000, y


download_assets()

title = "vietTTS"
description = "A vietnamese text-to-speech demo."

gr.Interface(
    fn=speak, 
    inputs="text", 
    outputs="audio",
    title = title,
    description=description,
    theme="default",
    allow_screenshot=False,
    allow_flagging="never",
).launch(debug=False)