Emotional Speaker Verification Based on I-vectors

Recently i-vectors approach in speaker verification become very successful and popular. The i-vectors principle is based on representation of each utterance by low-dimensional feature vector of fixed length. In this experiment for purposes of speaker recognition emotional speech database was applied. Using the i-vector principle two concepts of speaker model training were performed. In the process of features extraction the Mel Frequency Cepstral Coefficients (MFCC) with different number of coefficients in combination with coefficient of log energy, the first, second and third regression coefficients were used. Mahalanobis distance metric and Cosine Distance Scoring (CSS) metric were used for classification of the speaker recognition in this paper. In this work our own emotional database - SUS - of recordings in Slovak language was introduced. Utterances of male speakers of mentioned corpus were used as an input to the verification system.