Effects Of Background Data Duration On Speaker Verification Performance

Gaussian mixture models with universal background model (GMM-UBM) and vector quantization with universal background model (VQ-UBM) are the two well-known classifiers used for speaker verification. Generally, UBM is trained with many hours of speech from a large pool of different speakers. In this study, we analyze the effect of data duration used to train UBM on text-independent speaker verification performance using GMM-UBM and VQ-UBM modeling techniques. Experiments carried out NIST 2002 speaker recognition evaluation (SRE) corpus show that background data duration to train UBM has small impact on recognition performance for GMM-UBM and VQ-UBM classifiers.