Speaker verification based on the orthogonalisation technique

This paper describes an investigation into text-dependent speaker verification using orthogonal feature parameters of speech. The study is based on the use of a subset of the BT Millar speech database, consisting of repetitions of digit utterances 1 to 9 and zero spoken by 20 male speakers. Sets of orthogonal parameters are obtained through a linear transformation of linear prediction, parcor, and cepstrum coefficients. It is demonstrated that amongst these, orthogonal cepstrum parameters possess the highest speaker discrimination ability. Based on the experimental results it is shown that for single-digit inputs, a minimum equal error rate of about 4% in verification can be achieved. This error rate is found to reach 0.95% when a sequence of five digits is used as the input. The experiments are discussed in detail, and an analysis of the results is presented.