OPTIMAL UNIVERSAL BACKGROUND MODEL IN AUTOMATIC SPEAKER VERIFICATION

摘要

Automatic Speaker Verification (ASV) is a binary classification task to decide whether a claimed speaker uttered a sentence. This paper proposes two different algorithms for vector quantization (VQ) to speaker verification. Our first algorithm called Sub Vector Quantization (Sub VQ) is based on multiples codebook, represents the target speakers and the universal background model (impostors) and compared it to second vector quantization algorithm used for reducing training data. We compared our algorithms with the baseline system: Gaussian Mixtures Models and Maximum a Posteriori Adaptation. The present study demonstrates that the several codebooks for Universal Background Models give better results and less error rate. The performance of these models is evaluated on the Arabic speaker verification dataset. The VQ SUB method achieved less false acceptance and false reject error for 128 codebook size better than Baseline Vector Quantization approach for 32 codebook size. However, this improvement also depends on the codebook size.