Speaker Dependent Bengali Keyword Spotting in Unconstrained English Speech Acknowledgement

A project report submitted during summer internship under the supervision of Prof. Abstract Multi‐lingual interfaces can be of great use in a number of applications. A very important issue for such systems is to first identify the segments of utterances corresponding to a specific language. Language boundary information is also very vital before any further processing can be done. Language specific keyword spotting can be used for this purpose. Thus such a word spotter can serve as an integral part of a typical multi‐lingual system. A speaker dependent 'Bengali' keyword spotter in unconstrained 'English' speech had been developed in this project. Two approaches were used. Both used whole word based HMMs for keywords. All the Bengali keywords were trained as isolated words. The first approach used whole word filler model. The second approach used trained English phoneme models with an all phone grammar network to model the filler part. For whole word based approach an optimal performance of 94.22% hit with 1.17 FA/KW/H was obtained while the maximum %hit for the same system was 97.92% but at the cost of 7.03 FA/KW/H. The second approach attained an optimal performance with hit rate of 95.83% with just 0.71 FA/KW/H. However, maximum %hit for this system was same as first approach but with lesser false alarm rate of 4.45 FA/KW/H. Performance improvements in terms of reduction of false alarms have also been proposed. Finally, further development of the existing system to a 'speaker independent Bengali keyword spotter' has been discussed.