Journal of Intelligent Systems (Oct 2016)

An Experimental Comparison of Modeling Techniques and Combination of Speaker – Specific Information from Different Languages for Multilingual Speaker Identification

  • Jayanna H.S.,
  • Nagaraja B.G.

DOI
https://doi.org/10.1515/jisys-2014-0128
Journal volume & issue
Vol. 25, no. 4
pp. 529 – 538

Abstract

Read online

Most of the state-of-the-art speaker identification systems work on a monolingual (preferably English) scenario. Therefore, English-language autocratic countries can use the system efficiently for speaker recognition. However, there are many countries, including India, that are multilingual in nature. People in such countries have habituated to speak multiple languages. The existing speaker identification system may yield poor performance if a speaker’s train and test data are in different languages. Thus, developing a robust multilingual speaker identification system is an issue in many countries. In this work, an experimental evaluation of the modeling techniques, including self-organizing map (SOM), learning vector quantization (LVQ), and Gaussian mixture model-universal background model (GMM-UBM) classifiers for multilingual speaker identification, is presented. The monolingual and crosslingual speaker identification studies are conducted using 50 speakers of our own database. It is observed from the experimental results that the GMM-UBM classifier gives better identification performance than the SOM and LVQ classifiers. Furthermore, we propose a combination of speaker-specific information from different languages for crosslingual speaker identification, and it is observed that the combination feature gives better performance in all the crosslingual speaker identification experiments.

Keywords