Using Sub-Word Units for Low-Resource Language Keyword Searching
Tianqi Xu, Wei-Qiang Zhang, Yao Liu · 2020
The traditional pipelined keyword search system suffered from the out-of-vocabulary (OOV) keywords for low-resource languages. To investigate the performance of different sub-word units on OOV recovery, we make some experiments with BABEL Pashto dataset. The experiment results indicate the phoneme search method combined with grapheme-to-phoneme (G2P) converter gets the best actual term-weighted value (ATWV) score of all the sub-word units. In addition, to solve the high false alarm rate of the sub-word method, we employ the system combination method and get the comparable results to the proxy word method in Kaldi toolkit.