Do we need to customize characters for Arabic language?

[cod3r0k]

Hi do we need to edit config of characters as below:

# INITIALIZE THE TOKENIZER
# Tokenizer is used to convert text to sequences of token IDs.
# config is updated with the default characters if not defined in the config.
tokenizer, config = TTSTokenizer.init_from_config(config)
tokenizer.characters.punctuations = "!'(),-.:;? ،؟ ‌ "

[eginhard]

Yes, they should be updated for your target language, so that e.g. the model can learn the correct intonation for questions. Otherwise, any unknown characters will be removed from the input.

[cod3r0k]

if we add arabic characters in tokenizer.characters.punctuations, what is it happend? coqui remove them? or coqui try to learn and focus how they contribute in audio? how it does alignment? or..

[eginhard]

The preprocessing steps are roughly:

1. Input text: "<Hello, world!>"
2. Text cleaners (e.g. remove unwanted characters, lower case): "hello, world!"
3. Remove punctuation: "hello world"
4. Phonemize (for phoneme-based models): "həlˈoʊ wˈɜːld"
5. Restore punctuation: "həlˈoʊ, wˈɜːld!"

If the punctuation wasn't restored, the model couldn't learn differences between "hello!" and "hello?".

[cod3r0k]

Thank you so much, @eginhard! Do we need to process them manually before feeding them to Coqui, or should we add new functions in tts\utils\text\cleaners.py for Arabic in Coqui?
How can we handle the fifth level? Do you have any guidance in Coqui for that? I couldn’t find any details about restoring them. Does Coqui handle it?

[eginhard]

Yes, all the steps are already handled by Coqui. I don't know if a custom cleaner would make sense for Arabic, but probably starting with the basic_cleaner is fine.

[cod3r0k]

Great! Based on what you said, it is important to add a custom character as shown below for training:

1- First, initialize it:
tokenizer, config = TTSTokenizer.init_from_config(config)

2- Edit it
tokenizer.characters.punctuations = "!'(),-.:;? ،؟ ‌ "

About adding arabic_cleaner, I mean for example we can remove diacritics which we have in Arabic for better pronunciation or we can replace any specific words for abbreviation and so on @eginhard

[SyamsQ]

I need Arabic