Development of Multilingual Speech Database for Speaker Recognition in Indian Languages

Abstract

In this paper, we describe the collection of speech samples to develop a database for speaker recognition in the Indian scenario in the office environment and named VIT-Indian Language Speech Corpus (VIT-ILSC) speech database. Presently, we developed the Phase −1 database of speech samples from 50 speakers. The speech samples were collected in the office environment. Most of the speech samples collected are in English and other Indian languages in reading style, using two digital voice recorders. This work aims to develop a speech corpus database for a speaker recognition system in Indian languages, including English. Traditional Mel-frequency cepstral coefficients (MFCC) and Gaussian Mixture Model (GMM) was used to evaluate the collected phase-1 database. The phase-1 database has been evaluated on a speaker verification system. We considered both clean and noise backgrounds for initial studies and showed the impact of mismatch in training and testing samples.