SVM based speaker verification system using vowel like and non-vowel like regions

In this paper, speaker verification is done by using vowel like regions and non-vowel like regions. These VLRs and Non-VLRSs having different energies are determined from LP residual samples. By applying different normalization to these two regions the gross level mismatch and degradation effects are considerably reduced. The vowel like regions and non-vowel like regions are used independently for training and testing phases of SV systems. Vowel like regions is assigned higher weights due to their speaker specific information than non-vowel like regions. Finally the scores from VLRs and Non-VLRs are linearly combined to give the best match. This speaker verification process is carried out using TIMIT database, which gives a better performance in terms of equal error rate (EER) for a GMM-UBM based and an i-vector based SV system.