Multilingual keyword spotting is of immense interest in the Indian context with as many as 30 languages spoken across the country, and more than one language spoken in most cities. In this paper, we present the details of a gender independent multilingual keyword spotting system developed using lattices generated by a multilingual phone decoder for two of the most widely spoken Indian languages, Hindi and Tamil. For building the multilingual phone decoder, we used phonetic as well as acoustic similarities to map phones across the two languages, and see that the approach offers promising results. A distance measure based on Kullback-Leibler divergence is used for measuring the acoustic similarity of phones. We used a hybrid hidden Markov model – neural network implementation of the phone decoder for all our experiments reported in this work.
Dr. Santhosh Kumar C. and Mohandas, V. P., “Keyword Spotting in Multilingual Environments”, International Journal of Computer and Electrical Engineering, vol. 2, p. 1025, 2010.