Skip to content

NLP preprocessor

Jump to bottom
Anton edited this page Apr 9, 2021 · 7 revisions

Main page: sova-tts-tps

Info

This package was created in order to construct, on a modular basis, the text processors necessary to obtain a text prepared for submission to speech synthesis systems. At the moment, only Russian and English languages are supported. There is only basic support for English (see Basic functionality); following modules modules are implemented for Russian:

The project has the potential for expansion with other modules (g2p, normalizer) and languages.

How to use

Modules

Let us say you need to stress words in some text. You need stress setting module (Emphasizer) to complete this task:

from tps import download
from tps.data import find
from tps.modules import Emphasizer


try:
    stress_dict = find("stress.dict", raise_exception=True)
except FileNotFoundError:
    stress_dict = download("stress.dict")

emphasizer = Emphasizer("ru", (stress_dict, "plane"))
text = "Привет, мир! Смотри, как я умею ставить ударения в словах."
emphasizer.process_text(text)

results in

'Привет, м+ир! Смотри, как я ум+ею ставить удар+ения в слов+ах.'

Since this module is case sensitive, you get a different result if you lowercase the text:

emphasizer.process_text(text.lower())

results in

'прив+ет, м+ир! смотри, как я ум+ею ставить удар+ения в слов+ах.'

Handler

Let us say you need to prepare a text for transfer to the speech synthesis system. There are two possible scenarios for you.

Basic functionality

Let us say you already have a cleaned up and marked up text for training synthesis system. In this case, you only need the basic functionality of the Handler class, such as:

  • storing a fixed symbolic dictionary
  • conversion to lower case
  • obtaining vectors of sentences based on a symbolic dictionary
  • getting suggestions from vectors
  • some additional cleaning

In that case the following code is enough

from tps import Handler

handler = Handler("ru")
text = "В чащах юга жил бы цитрус? Да, но фальшивый экземпляр!"

result = handler.process_text(text, keep_delimiters=False)
print(result)
'в чащах юга жил бы цитрус? да, но фальшивый экземпляр!'

vector = handler.text2vec(result)
print(vector)
[16, 12, 38, 14, 40, 14, 36, 12, 45, 17, 14, 12, 21, 23, 26, 12, 15, 42, 12, 37, 23, 33, 31, 34, 32, 4, 12, 18, 14, 9, 12, 28, 29, 12, 35, 14, 26, 43, 39, 23, 16, 42, 24, 12, 44, 25, 22, 19, 27, 30, 26, 46, 31, 3]

Connecting modules

In case you need to prepare text for inference, you may need to connect various modules that improve the text sent for synthesis. For Russian language just use from_charset method of Handler class (silent=True allows loading in automatic mode the data required for the modules to work):

from tps import Handler

handler = Handler.from_charset("ru", silent=True)
text = "В чащах юга жил бы цитрус? Да, но фальшивый экземпляр!"

result = handler.process_text(text, keep_delimiters=False)
print(result)
в ч+ащах +юга ж+ил бы ц+итрус? да, но фальш+ивый экземпл+яр!

or, if you need to preserve delimiters:

result = handler.process_text(text, keep_delimiters=True)
print(result)
['в ч+ащах +юга ж+ил бы ц+итрус?', <Pause.eos: 500ms>, 'да, но фальш+ивый экземпл+яр!']

also there is possibility to pass the user dictionary and necessary cleaner functions (including custom ones)

text = "TTS     -    это     увлекательно."
user_dict = {"tts": "синтез речи"}

result = handler.process_text(text, cleaners="light_punctuation_cleaners", user_dict=user_dict, keep_delimiters=False)
'с+интэз р+ечи — +это увлек+ательно.'

In case you want to link your other modules with the handler, then do this when initializing the class instance

handler = Handler("ru", modules=some_modules_list)

How to add new module

How to add new language

Clone this wiki locally