Title:
Combining Different Lexical Units in RNN-based Language Modeling for Open-domain Thai Speech Recognition

Speaker:
WUTIWIWATCHAI, Chai (NECTEC)

Abstract:
Open-domain speech recognition requires a large coverage of vocabulary in the language. While it is not possible to include all words in the language in the system dictionary, a general practice is to break words down to sub-word units which can efficiently help increase the vocabulary coverage. While smaller units like sub-words have a drawback of acoustic confusing, a compromising method is to include both words and sub-words in the dictionary. In Thai, a sub-word unit is defined as a pseudo-morpheme (PM) which is a syllable-like unit. Frequently used PM units combined with frequently used word units have been combined and showed effectiveness in open-domain Thai speech recognition both using weighed finite-state transducer (WFST) and recently introduced recurrent neural network (RNN) based language modeling. This talk summarizes our practice in combining such different lexical units and discusses on issues potentially explored to further improve the system performance.