A Feature Selection based Approach for Speaker and Language Identification in VoIP Networks

fulfillment for the award of the degree of Master of Technology (Computer and Information Technology) is a bona-fide record of work carried out by him under our supervision and guidance. The thesis has fulfilled all the requirements as per the regulations of the Institute and, in our opinion, has reached the standard needed for submission. ii Acknowledgements The writing of this thesis has been a long and arduous journey of learning. From the initial concept to the final product, this thesis has been the input from many individuals. I extend my most sincere gratitude to Prof. Indranil Sengupta and Dr. S.K. Ghosh, my supervisors, for their continued support, constructive comments and encouragement. Without their excellent guidance, this work would not have taken the present shape. I feel privileged to have got this opportunity to work with them and learn from their vast academic and research experience. and Political Science for sharing his knowledge and valuable thoughts for the improvements in the basic approach presented in this thesis. Also I would like to Dr. M. Patracca of Politecnico di Torino, Italy for suggesting valuable improvements in few experiments presented in this thesis. Lastly and most importantly, I am thankful and indebted to my organization, the Indian Army, for giving me an opportunity to pursue higher education and research in an elite academic institute. I am grateful for the love and support of my family – my father and my wife, Sukhjeet and my two sons Maitrey and Tej for their love and affection, which kept me motivated throughout my work on this thesis. Without their support, I would not have been able to make it this far. Abstract In VoIP networks, generally the speech is transmitted in the compressed format using some speech compression algorithm, whereas typical automatic speaker or language identification systems are not capable of handling compressed speech. Hence the compressed speech has to be re-synthesized to original waveform in order to get it processed through a speaker or language identification system which involves a considerable computational overhead. Thereby the speaker and language identification systems based on normal speech waveform are not suitable for most of the futuristic speaker or language identification based VoIP applications where the ability to identify a number of speaker or languages in real-time will be essential. In this thesis, we propose an on-line speaker and language identification scheme where Dimensionally Reduced Significant Statistical (DRSS) …