tihu-nlp / tihudict Goto Github PK

Thanks Mostafa for the repository. NLP and speech-related products are so poor for Persian language. We need to survive it.
The pre-trained model in the release page is not compatible with the referred source code provided in https://github.com/cmusphinx/g2p-seq2seq. I have tried to load the model with the source code but the checkpoint is missing many parameters such as "global_step."
As the pre-trained model without the code company is useless, I appreciate you to provide it.

Obvious care has been taken to make sure the pronunciation of ه (hā-ye do-češm) has been differentiated between its pronunciation as /h/ (word-initially and -medially) and /e/ (word-finally) as in:

• ماهه m A h e

Examples like that above are useful to make sure when the grapheme should be pronounced as [h] or [e]. However, there does not seem to be a distinction between the [h] pronunciation of ه (hā-ye do-češm) and the [h] pronunciation of ح (ḥâ-ye ḥotti / ḥâ-ye jimi) anywhere in the dictionary. Consider the following examples:

• [1] حادثه h A d e s e
• [2] حوزه h o z e
• [3] صحیح s a h i h
• [4] آنها A n h A

If this dictionary were to be used in its reverse form, [1] could be reconstituted from "h A d e s e" into either "هادِثه" or "حادِثه". This is certainly a niche issue, but I am trying to diacritize non-diacritized text, and so in order to re-constitute the original text I have with included vowels given your dictionary's scheme, I need to know which Farsi character to convert back to in the end. There are no instances in either dictionary where ḥâ-ye jimi and hā-ye do-češm (pronounced as [h]) appear in the same word so a simple string replace should do it there. I suggest replacing [h] with [H] for
ḥâ-ye jimi to make this dictionary reversible.

It may well be that there are no words in Farsi which contain both ح and ه, but this would solve the edge use-case I describe here.

Thanks for your work!

I need something like this for studying, but larger