On the Use of Gaussian Mixture Model (GMM) Technique and YOHO Corpora for Automatic Speaker Recognition for Nigerian Tribal Languages

Many levels of information can be gotten from speech such as the message being spoken, the language been spoken, information about the speaker and the emotional state of the speaker. Several literature have reported on automatic speaker recognition system. This article overviewed speaker recognition systems with emphasis on those using Gaussian Mixture Model (GMM). Subsequently, via a statistical based speaker-modeling technique that represents the underlying characteristic sounds of a person's voice an analysis of the speech of speakers from Hausa, Yoruba and Igbo tribes in Nigeria was performed. Speaker recognizers that are capable of recognizing a speaker that is text-independent was designed. The wavelet toolbox of MATLAB® 2007 was used. Performance of the systems is evaluated for a wide range of speech quality; from clean speech to cell-phone speech, by using YOHO standard speech corpora. An Error Rate (ERR) of false acceptance rate of 0.51% and a false rejection rate of 0.65% at a 0.1% false-acceptance rate were obtained. ERR of compensation channel of 0.96%. Same experiment in identification section was also conducted, consequently 0.28% identification error rate was achieved. Using 4.45% of the speaker utterance 0.7% identification error rate was achieve.