MahmoudAshraf97 / ctc-forced-aligner

Text to speech alignment using CTC forced alignment
130 stars 26 forks source link

AssertionError with blank side #17

Closed rehandaphedar closed 1 month ago

rehandaphedar commented 1 month ago

I'm getting the same error as #16 again. This time, it reads:

AssertionError: ا != 

I see that the previous error was:

AssertionError: ن != ٰ

You added ٰ to the normalisation config. I thought of making a similar PR, but this time there is nothing on the other side.

text.txt audio.wav

MahmoudAshraf97 commented 1 month ago

these errors will continue to appear with arabic, can you test as much audios as possible so we can improve the normalization? each new error will create a new normalization rule

rehandaphedar commented 1 month ago

Sure, I wanna try to create word timings for the complete Qurʾān, so that should be plenty of testing data.

MahmoudAshraf97 commented 1 month ago

the included text does not match the audio

rehandaphedar commented 1 month ago

Sorry, I posted the wrong text.txt link. This is the correct text.txt.

MahmoudAshraf97 commented 1 month ago

I fixed it and some more, please note that this approach is not the correct approach, Quran normalization is not a simple search and replace, it needs more hand-crafted rules that I will not be covering in this repo, so if you want to keep using an Arabic model please handle these rules yourself using the previous two commits as a guide, or use the default model with romanization which will give you better results

rehandaphedar commented 1 month ago

Noted.