Optimal impostor model in automatic speaker verification

Abstract

Speaker Verification (ASV) is a binary classification task to decide whether a claimed speaker uttered sentences. This paper proposes two different algorithms for vector quantization (VQ) to speaker verification. The first algorithm named Partial Vector Quantization (Partial VQ) is based on partitioning acoustics space, represents the impostors called universal background model(UBM) and compared it to second vector quantization algorithm Reduced UBM session used for keeping the relevant training data. The present study demonstrates that several codebooks for Universal Background Models give better results. The performance of these models is evaluated on the Arabic speaker verification dataset. The VQ Partial method achieved less half total error rate for 128 codebook size better than Baseline Vector Quantization approach for 32 codebook sizes.