Why am I interested in this? These guys agreed to make several new synthesizers for Uzbek, Tajik, Persian and some other languages.
I tried to run it for Windows, and it was very, very difficult!
The question is, is it possible to create a NVDA or SAPPY synthesizer from this?
Dear developers, please look at it and comment. I like the quality of voices, but it's possible to run only from the code!

Here is a simple example. The only problem, Audio is an array of audio data to play or write to the file. While I don't know how.
Attention! Here is Russian text, but it is possible to use an English model with English text.

import os
import torch
from utils import *

device = torch.device('cpu')
torch.set_grad_enabled(False)
symbols = '_~абвгдеёжзийклмнопрстуфхцчшщъыьэюя +.,!?…:;–'
local_file = 'model.jit'

if not os.path.isfile(local_file):
  torch.hub.download_url_to_file('https://models.silero.ai/models/tts/ru/v1_kseniya_16000.jit',
                                 local_file)

model = torch.jit.load('model.jit',
                       map_location=device)
model.eval()
example_text = 'Это просто пример!'
sample_rate = 16000
model = model.to(device)  # gpu or cpu

audio = apply_tts(texts=[example_text],
                  model=model,
                  sample_rate=sample_rate,
                  symbols=symbols,
                  device=device)

print(example_text)
display(Audio(audio[0], rate=sample_rate))

@camlorn
Not fast, Although the Creator assures that it's fast.
But, most importantly, I would like to get the opportunity to create audio books using these synthesizers.
There are some problems, for example, it reads only one sentence
to 140 characters.
I would like to know about creating it for SAPPY-5 to create audio books.

You're probably not going to get what you want out of this.

Time your script.  If it's slower than 200 milliseconds it won't work in a screen reader.

Most of these speech models have a character limit. You'd need to split the text into sentences first, but many sentences are longer than 140 characters.

Machine learning text to speech systems are still kind of far away when it comes to running on your computer as part of sapi or something.

@9
If you want audiobooks you don't need sapi and can just run the algorithm directly on the text files.

If the script takes 1 second per sentence then it takes 1 second before you can start listening at all.  If you time how long your script takes to run and tell us, we can tell you what you might be able to do with it.  I'm not sure why you're not just timing it given that you've already got one?  Are you unable to run it?

You can write a convenient program to make audiobooks with speech synthesis in like 50 lines of Python tops. Sapi will take 500 or more.  In any case the character limit means that you will have to do some clever stuff to even make it work asuming that you can in the first place.

We don't care how long it takes to write to a file. Time this line:

audio = apply_tts(texts=[example_text],
                  model=model,
                  sample_rate=sample_rate,
                  symbols=symbols,
                  device=device)

@jonikster,
many of the deep learning models, are even have a lot of computations during inference (the computations are worse during training always).
now, about making audio version of books, you need to present those texts to the model (as it should be less than or equal to the sequence length of the model,which probably exceeds it, if it is tacotron or a rnn based model).
if it is a convolutional only based model (like DCTTS), that will be a little different even in terms of speed, and sequence generations (which text and audio are), which I assume, these models are not.

@15
No, probably not.  That was the point I was trying to make.  Jonikster has seen their "in realtime on a CPU" claim and has drawn the wrong conclusion from it.  This stuff is actually that slow.

You might be able to get a really fast version of it running if you have a compatible GPU and jump through a bunch of hoops to get whatever ml framework they're using to recognize it, and maybe that'd work for audiobooks in a reasonable amount of time.  Honestly 1 word every 5 seconds works for audiobooks if you don't mind running it for days, so it could still maybe work for that.

I see the point of this: OP speaks a language without good synths.  This is part of why I want to write one, because I think that properly applied and sacrificing characteristics that sighted people care about, it could be pretty automatic.  But you really can't just shove these models into a laptop.  The unspoken caveat about that speed thing is that they're measuring it on really beefy servers with more cores than my tower and a bunch of SSDs and stuff.

@17, there already is one