Build NEW Acoustic model, Dictionary , Language model for uncommon language speech recognition

Question

I want to build NEW Acoustic model ,New Dictionary ,New Language model for "Sinhala Language speech recognition" Sinhala language Characters are Unicode based. for an example A=අ,I=ඉ,U=උ,KA=ක,BA=බ. I did go through CMUSphinx Tutorial For Developers. But it did not help me. It works for English language.

Language model should be ARPA model. and How can I map Sinhala Unicode with English phonemes and how to train Language model with Different voices. Is there any tool available for generate Unicode based language model?

Answer 1

Overall, it is not really complex. First you need to split the task on parts: build phonetic dictionary, build language model, build acoustic model. Start with phonetic dictionary.

You need to write a Python script to map unicode input to the transliteration:

රට  r a tt a
එකඟයි   e k a ng a yi
අවසර දිම    a v a s a r a d i m a

Basically for every you write a corresponding transliteration. That is all you need to do, later you can just feed the list of words into your script and get a dictionary in cmusphinx format. This part is covered in tutorial

http://cmusphinx.sourceforge.net/wiki/tutorialdict

Once you have a transliteration tool you can proceed with language model. You need a lot of texts to build a language model. You can download texts from wikipedia or from local newspaper. Then you can use any language model toolkit to create an ARPA model. All of them support unicode - SRILM, MITLM, IRSTLM, you can use any of them. This part is covered in tutorial

http://cmusphinx.sourceforge.net/wiki/tutoriallm

Third step is to create an acoustic model. You need to record audio or segment existing recordings and start training. This part is also covered in the tutorial

http://cmusphinx.sourceforge.net/wiki/tutorialam