dataset

[VafaKnm]

Hi!
I have a question about dataset. Suppose that I have several wavs and the corresponding text files of them (written in Persian language). How I can create phoneme_transcriptions of them?

[Adibian]

Hi!
Based on my experiments, for training any of TTS model for Persian language you need audio files and their phoneme sequences. You can not use raw Persian text and if you do, the result will not be good because of not written short vowels and Kasre_Ezafe.
Also creating phonemes from Persian text is not a simple task because it needs large lexicon, Grapheme_to_Phoneme model (G2P)(for words do not exist in lexicon), Ezafe prediction, and word sense disambiguation model (for words with multiple phonemes like 'mard' and 'mord').
And I don't know if there is any public tool or repository that handel all this problems and create phonemes from text or not.

[VafaKnm]

Thanks for sharing your experiences my friend.
Actually, I find a G2P model for Persian language:
https://github.com/PasaOpasen/PersianG2P
It's good but not perfect; for example it can't recongnize Kasre between two words ("gol ziba" instead of "gole ziba") but anyway I have no other choice!

I have one more question; Why there is not any spaces between the words? for example what happen if we build dataset like "gole ziba" instead of "goleziba"?

[Adibian]

In the speech synthesis from the phoneme sequence, space is not important. Because you have to separate the phonemes, consider the ID number for each phoneme, and use the sequence of IDs, so the spaces between the words are removed in this process. Of course, you can consider a new token (like other phonemes) for the space between words but note that usually the duration of this phoneme will be zero or very little.