app.py

import os
import sys
import importlib
from subprocess import call
from pathlib import Path

import json
import math
import matplotlib.pyplot as plt
import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import DataLoader
from scipy.io.wavfile import write
import gradio as gr
import scipy.io.wavfile
import numpy as np
from libs.audio import wav2ogg, float2pcm

def run_cmd(command):
    try:
        # print(command)
        call(command, shell=True)
    except KeyboardInterrupt:
        print("Process interrupted")
        sys.exit(1)

current = os.getcwd()
full = current + "/vits/monotonic_align"
os.chdir(full)
run_cmd("python3 setup.py build_ext --inplace")
run_cmd("mv vits/monotonic_align/* ./")
run_cmd("rm -rf vits")
# run_cmd(f"mv {current}/lfs/*.pth {current}/pretrained/")
# run_cmd("apt-get install espeak -y")
# run_cmd("gdown 'https://drive.google.com/uc?id=1q86w74Ygw2hNzYP9cWkeClGT5X25PvBT'")
os.chdir("../..")

from lojban.lojban import lojban2ipa

sys.path.insert(0, './vits')
import vits.commons as commons
import vits.utils as utils
from vits.data_utils import TextAudioLoader, TextAudioCollate, TextAudioSpeakerLoader, TextAudioSpeakerCollate
from vits.models import SynthesizerTrn
from vits.text.symbols import symbols
from vits.text import _clean_text
from vits.text import cleaners
from vits.text.symbols import symbols

sys.path.insert(0, './nix_tts_simple')
from nix_tts_simple.tts import generate_voice

language_id_lookup = {
    "Lojban": "jbo",
    "Transcription": "ipa",
    "English": "en",
    "German": "de",
    "Greek": "el",
    "Spanish": "es",
    "Finnish": "fi",
    "Russian": "ru",
    "Hungarian": "hu",
    "Dutch": "nl",
    "French": "fr",
    'Polish': "pl",
    'Portuguese': "pt",
    'Italian': "it",
}

# def download_pretrained():
#     if not all(Path(file).exists() for file in ["pretrained_ljs.pth", "pretrained_vctk.pth"]):
#         url = "https://drive.google.com/uc?id=1q86w74Ygw2hNzYP9cWkeClGT5X25PvBT"
#         gdown.download_folder(url, quiet=True, use_cookies=False)

# Mappings from symbol to numeric ID and vice versa:
_symbol_to_id = {s: i for i, s in enumerate(symbols)}
_id_to_symbol = {i: s for i, s in enumerate(symbols)}


def text_to_sequence(text, language, cleaner_names, mode):
    '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
        Args:
        text: string to convert to a sequence
        cleaner_names: names of the cleaner functions to run the text through
        Returns:
        List of integers corresponding to the symbols in the text
    '''
    sequence = []

    if language == 'jbo':
        clean_text = lojban2ipa(text, mode)
    elif language == 'ipa':
        clean_text = text
    else:
        clean_text = _clean_text(text, cleaner_names)

    for symbol in clean_text:
        symbol_id = _symbol_to_id[symbol]
        sequence += [symbol_id]
    return clean_text, sequence


def get_text(text, language, hps, mode):
    ipa_text, text_norm = text_to_sequence(
        text, language, hps.data.text_cleaners, mode)
    if hps.data.add_blank:
        text_norm = commons.intersperse(text_norm, 0)
    text_norm_tensor = torch.LongTensor(text_norm)
    return ipa_text, text_norm_tensor


def load_checkpoints():
    hps = utils.get_hparams_from_file(current + "/vits/configs/ljs_base.json")
    model = SynthesizerTrn(
        len(symbols),
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        **hps.model)
    _ = model.eval()

    _ = utils.load_checkpoint(current + "/pretrained/vits/pretrained_ljs.pth", model, None)

    hps_vctk = utils.get_hparams_from_file(current + "/vits/configs/vctk_base.json")
    net_g_vctk = SynthesizerTrn(
        len(symbols),
        hps_vctk.data.filter_length // 2 + 1,
        hps_vctk.train.segment_size // hps_vctk.data.hop_length,
        n_speakers=hps_vctk.data.n_speakers,
        **hps_vctk.model)
    _ = model.eval()

    _ = utils.load_checkpoint(current + "/pretrained/vits/pretrained_vctk.pth", net_g_vctk, None)

    return model, hps, net_g_vctk, hps_vctk

def inference(text, language, noise_scale, noise_scale_w, length_scale, voice, file_format):
    if len(text.strip())==0:
        return []
    language = language.split()[0]
    language = language_id_lookup[language] if bool(
        language_id_lookup[language]) else "jbo"
    result = []
    if voice == 'Nix-Deterministic' and language == 'jbo':
        result = generate_voice(lojban2ipa(text,'nix'), current+"/pretrained/nix-tts/nix-ljspeech-v0.1")
    elif voice == 'Nix-Stochastic' and language == 'jbo':
        result = generate_voice(lojban2ipa(text,'nix'), current+"/pretrained/nix-tts/nix-ljspeech-sdp-v0.1")
    elif voice == 'LJS':
        ipa_text, stn_tst = get_text(text, language, hps, mode="VITS")
        with torch.no_grad():
            x_tst = stn_tst.unsqueeze(0)
            x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
            audio = model.infer(x_tst, x_tst_lengths, noise_scale=noise_scale,
                                noise_scale_w=noise_scale_w, length_scale=length_scale)[0][0, 0].data.float().numpy()
            result = [ipa_text, (hps_vctk.data.sampling_rate, audio)]
    else:
        ipa_text, stn_tst = get_text(text, language, hps_vctk, mode="VITS")
        with torch.no_grad():
            x_tst = stn_tst.unsqueeze(0)
            x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
            sid = torch.LongTensor([voice])
            audio = model_vctk.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=noise_scale,
                                     noise_scale_w=noise_scale_w, length_scale=length_scale)[0][0, 0].data.cpu().float().numpy()
            result = [ipa_text, (hps_vctk.data.sampling_rate, audio)]
    if file_format == 'ogg':
        result = [result[0], wav2ogg(result[1][1], result[1][0], text, language)]
    else:
        result = [result[0], (result[1][0], float2pcm(result[1][1]))]

    return result

# download_pretrained()
model, hps, model_vctk, hps_vctk = load_checkpoints()

defaults = {
    "text": "coi munje",
    "language": "Lojban",
    "noise_scale": .667,
    "noise_scale_w": .8,
    "speed": 1.8,
    "voice": "LJS",
    "example": ["", "Lojban", 0.667, 0.8, 1.8,"LJS","wav"]
}

inputs = []
outputs = []

css = """
h1 {font-size:200%;}
h2 {font-size:120%;}
h2 a {color: #0020c5;text-decoration: underline;}
p a {text-decoration: underline;}
img {display: inline-block;height:32px;}
#velsku {
  text-align: left;
  margin: 0;
  /* display: none; */
  /* justify-content: baseline; */
  /* align-items: flex-start; */
  padding: 0;
  /* height: 28px; */
  width: 100%;
  bottom: 0px;
  left: 0px;
  position: fixed;
  background: white;
  z-index: 10;
}

#velsku_sebenji {
  padding: 0.1rem;
  display: flex;
  white-space: nowrap;
  text-overflow: ellipsis;
  box-shadow: 0 0 0 1px rgb(56 136 233), 0 0 0 2px rgb(34 87 213),
    0 0 0 3px rgb(38 99 224), 0 0 0 4px rgb(25 65 165);
  margin-top: 3px;
}

.velsku_pamei {
  white-space: nowrap;
  overflow: hidden;
  text-overflow: ellipsis;
}

.velsku_pixra {
  max-height: 20px;
  margin-right: 0.2rem;
}

#velsku a {
  color: #2b79e0;
  text-decoration: none;
}
"""

def conditionally_hide_widgets(voice):
    if str(voice).startswith("Nix"):
        return gr.update(visible=False), gr.update(visible=False), gr.update(visible=False)
    else:
        return gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)

title = "<h1>la vitci voksa - <i><img src='/file/assets/jbolanci.png'/>Lojban text-to-speech</i></h1>"
description = "<h2>VITS & Nix-TTS text-to-speech adapted to Lojban. Join <a href='https://discord.gg/BVm4EYR'>Lojban Discord live chat</a> to discuss further.</h2>"
article = "<p style='text-align: center'><a href='https://github.com/jaywalnut310/vits'>VITS: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech</a> | <a href='https://github.com/rendchevi/nix-tts'>Nix-TTS: Lightweight and End-to-end Text-to-Speech via Module-wise Distillation</a></p>"
chat = """
<script src="https://cdnjs.cloudflare.com/ajax/libs/socket.io/4.5.1/socket.io.js"></script>"
<script>
      document.addEventListener("DOMContentLoaded", () => {
        var socket1Chat_connected;
        var socket1Chat = io("wss://jbotcan.org:9091", {
          transports: ["polling", "websocket"],
        });
        // if (socket1Chat) {
        socket1Chat.on("connect", function () {
          console.log(socket1Chat);
          socket1Chat_connected = true;
        });
        socket1Chat.on("connect_error", function () {
          console.log("1chat connection error");
        });

        function trimSocketChunk(text) {
          return text
            .replace(/[\n\r]+$/gims, " ")
            .replace(/<br *\/?>/gims, " ");
          // .split('<')[0]
        }
        socket1Chat.on("sentFrom", function (data) {
          if (!socket1Chat_connected) return;
          const i = data.data;
          const msg = {
            d: trimSocketChunk(i.chunk),
            s: i.channelId,
            w: i.author,
          };

          const velsku = document.getElementById("velsku_sebenji");
          velsku.innerHTML = `<img src="https://la-lojban.github.io/sutysisku/pixra/nunsku.svg" class="velsku_pixra"/> <span class="velsku_pamei">[${msg.s}] ${msg.w}: ${msg.d}</span>`;
        });
        socket1Chat.on("history", function (data) {
          if (!socket1Chat_connected) return;
          const i = data.slice(-1)[0];
          if (!i) return;
          const msg = {
            d: trimSocketChunk(i.chunk),
            s: i.channelId,
            w: i.author,
          };
          const velsku = document.getElementById("velsku_sebenji");
          velsku.innerHTML = `<img src="https://la-lojban.github.io/sutysisku/pixra/nunsku.svg" class="velsku_pixra"/> <span class="velsku_pamei">[${msg.s}] ${msg.w}: ${msg.d}</span>`;
        });
        // }
      });
    </script>
    <div id="velsku" class="noselect">
                <a id="velsku_sebenji" href="https://discord.gg/4KhzRzpmVr" target="_blank">
                <img src="https://la-lojban.github.io/sutysisku/pixra/nunsku.svg" class="velsku_pixra"/>
                Live chat on Discord
                </a>
                </div>
"""
with gr.Blocks(css=css) as demo:
    gr.HTML(title)
    gr.HTML(description)    
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(lines=4, value=defaults["text"], label="Input text", placeholder="add your text, or click one of the examples to load them")
            langs = gr.Radio([
                'Lojban',
                'English',
                'Transcription',
                ], value=defaults["language"], label="Language")
            voices = gr.Radio(["LJS", 0, 1, 2, 3, 4, "Nix-Deterministic", "Nix-Stochastic"], value=defaults["voice"], label="Voice")
            noise_scale = gr.Slider(label="Noise scale", minimum=0, maximum=2,
                step=0.1, value=defaults["noise_scale"])
            noise_scale_w = gr.Slider(label="Noise scale W", minimum=0, maximum=2,
                step=0.1, value=defaults["noise_scale_w"])
            slowness = gr.Slider(label="Slowness", minimum=0.1, maximum=3,
                step=0.1, value=defaults["speed"])
            file_format = gr.Radio(["wav", "ogg"], value="wav", label="File format")
            
            inputs = [input_text, langs, noise_scale, noise_scale_w, slowness, voices, file_format]
            
            # events
            vits_inputs = [noise_scale, noise_scale_w, slowness]
            voices.change(fn=conditionally_hide_widgets, inputs=voices,outputs=vits_inputs)

        with gr.Column():
            ipa_block = gr.Textbox(label="International Phonetic Alphabet")
            audio = gr.Audio(type="numpy", label="Output audio")        

            outputs = [ ipa_block, audio ]
            
            btn = gr.Button("Vocalize")
            btn.click(fn=inference, inputs=inputs, outputs=outputs, api_name="cupra")
            
            examples = list(map(lambda el: el[0:len(el)] + defaults["example"][len(el):], [
                ["coi ro do ma nuzba", "Lojban"],
                ["mi djica lo nu do zvati ti", "Lojban", 0.667, 0.8, 1.8,4],
                ["mu xagji sofybakni cu zvati le purdi", "Lojban", 0.667, 0.8, 1.8, "Nix-Deterministic"],
                ["ni'o le pa tirxu be me'e zo .teris. pu ki kansa le za'u pendo be le nei le ka xabju le foldi be loi spati", "Lojban"],
                [", miː dʒˈiːʃaː loːnʊuː doː zvˈaːtiː tiː.", "Transcription"],
                ["We propose VITS, Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech.", "English"],
            ]))
            gr.Examples(examples, inputs, fn=inference, outputs=outputs, cache_examples=True, run_on_click=True)
    gr.HTML(article)
    gr.HTML(chat)

demo.launch(server_name="0.0.0.0")