Research on Mongolian Spoken Term Detection Method Based on Segmentation Recognition
Bao Fei-lon · 2013
Mongolian is an agglutinative language.This special formation rule results in an amount of probably millions of words which are far beyond the coverage of the pronunciation dictionary of any current Mongolian Large Vocabulary Continuous Speech Recognition(LVCSR)System.On the other hand,even if the pronunciation dictionary is large enough to cover most of the Mongolian words,the recognition system still won't perform well due to the sparseness of training data.To avoid the poor coverage problem of pronunciation Dictionary,we proposed a segmentation-based LVCSR approach and trained its acoustic model and language model.This approach is integrated into Mongolian Spoken Term Detection(STD)system and tested on Mongolian speech data.Experimental results show that our segmentation-based LVCSR approach can recognize most of the Mongolian words successfully and both the precision and recall of the Mongolian STD system are greatly improved by converting most of the out-of-vocabulary words into their in-vocabulary form.