Speaker Identification in Spoken Language Mismatch Condition: An Experimental Study

Abstract

This paper describes the impact of spoken language variation in a multilingual speaker identification (SID) system. The development of speech technology applications in low resource languages (LRL) is challenging due to the unavailability of proper speech corpus. This paper illustrates an experimental study of SID on Eastern and Northeastern (E&NE) Indian languages in language mismatch conditions. For this purpose, several experiments are carried out using the LRL data to build speaker identification models. Here, spectral features are explored for investigating the presence of speaker-specific information. Mel frequency cepstral coefficients (MFCCs) and shifted delta cepstral (SDC) are used for representing the spectral information. Gaussian mixture model (GMM) and support vector machine (SVM)-based models are developed to represent the speaker-specific information captured through the spectral features. Apart from that, to build the modern SID i-vectors, time delay neural networks (TDNN), and recurrent neural network with long short-term memory (LSTM-RNN) have been considered. For the evaluation, equal error rate (EER) has been used as a performance matrix of the SID system. Performances of the developed systems are analyzed with native and non-native corpus in terms of speaker identification (SID) accuracy. The best SID performances are observed to be EER 10.52% after the corpus fusion mechanism.