VQ/GMM-Based Speaker Identification with Emphasis on Language Dependency

Abstract

The biometric recognition of human through the speech signal is known as automatic speaker recognition (ASR) or voice biometric recognition. Plenty of acoustic features have been used in ASR so far, but among them Mel-frequency cepstral coefficients (MFCCs) and Gammatone frequency cepstral coefficients (GFCCs) are popularly used. To make ASR language and channel independent (if training and testing microphones and languages are not same), i-Vector feature and unwanted variability compensation techniques like linear discriminative analysis (LDA) or probabilistic LDA (PLDA), within-class covariance normalization (WCCN) are extensively used. At the very present days, the techniques for modeling/classification that are used are Gaussian mixture models (GMMs), vector quantization (VQ), hidden Markov model (HMM), deep neural network (DNN), and artificial neural network (ANN). Sometimes, model-domain normalization techniques are used to compensate unwanted variability due to language and channel mismatch in training and testing data. In the present paper, we have used maximum log-likelihood (MLL) to evaluate the performance of ASR on the databases(DBs), namely ELSDSR, Hyke-2011, and IITG-MV SR Phase-I & II, based on MFCCs and VQ/GMM where the scoring technique MLL is used for the recognition of speakers. The experiment is carried out to examine the language dependency and environmental mismatch between training and testing data.