Turkish text-dependent speaker verification using i-vector/PLDA approach

摘要

I-vector feature extraction is the state-of-the-art technique for text-independent speaker recognition. There exist studies in literature utilizing i-vector approach for text-dependent speaker verification. However, its performance for Turkish speaker recognition remains unknown. In this study, the performance of i-vector approach is analysed on Turkish text-dependent speaker recognition database consisting of 59 speakers. Experimental results show that, traditional Mel-frequency cepstral coefficients modelled with Gaussian mixture model - universal background model (GMM-UBM) outperforms i-vector system. It is also observed that probabilistic linear discriminant analysis (PLDA) classifier using i-vector features does not bring any performance improvement over the standard cosine distance scoring (CDS) for Turkish text-dependent speaker verification.