Capturing Indian phonemic diversity with multiple posteriorgrams for Multilingual Query-by-Example Spoken Term Detection
Abhimanyu Popli, Arun Kumar · 2017
This work presents a novel technique to represent Indian phonemes using English articulatory motivated posteriorgrams. Changes are done in articulatory posteriorgrams trained in English language to bring out minute details which are significant in multilingual scenario. With these enhancements, we show that articulatory posteriorgrams can give comparable or slightly better performance than the state-of-the-art phonemic posteriorgrams in the context of three Indian languages viz. Hindi, Telugu and Bangla. In addition, we also show that the combination of phonemic and articulatory posteriorgrams trained on English language can perform between 4.6 to 5.9 % (absolute) better than phonemic posteriorgrams alone for the task of Multilingual Query-by-Example Spoken Term Detection without using diversity of any other language or acoustic features. This establishes the complimentary nature of articulatory and phonemic posteriorgrams.