Speaker Recognition from Emotional Speech Using I-vector Approach

Abstract

In recent years the concept of i-vectors become very popular and successful in the field of the speaker verification. The basic principle of i-vectors is that each utterance is represented by fixed-length feature vector of low-dimension. In the literature for purpose of speaker verification various recordings obtained from telephones or microphones were used. The aim of this experiment was to perform speaker verification using speaker model trained with emotional recordings on i-vector basis. The Mel Frequency Cepstral Coefficients (MFCC), log energy, their deltas and acceleration coefficients were used in process of features extraction. As the classification methods of the verification system Mahalanobis distance metric in combination with Eigen Factor Radial normalization was used and in the second approach Cosine Distance Scoring (CSS) metric with Within-class Covariance Normalization as a channel compensation was employed. This verification system used emotional recordings of male subjects from freely available German emotional database (Emo-DB).

Keywords

• speaker verification
• i-vectors
• total variability
• emotions