Evaluation of the generative and discriminative text-independent speaker verification approaches on handheld devices

This paper takes advantage of the “MIT Mobile Device Speaker Verification Corpus” (MIT-MDSVC) availability in order to evaluate the performance of three well known text-independent speaker verification approaches on handheld devices, considering the MIT-MDSVC as a representative corpus designed for robust speaker verification tasks on limited vocabulary and limited amount of training data collected on handheld devices. Several experiments with either mismatched testing conditions, or with samples collected from multiple test conditions were conducted for evaluating both text-independent approaches: generative (based on Gaussian Mixture Models) and discriminative (based on Support Vector Machines with Fisher kernel and GMM Supervector Linear kernel), without using the transcription of the utterances or knowledge about the acoustic conditions of the recordings (environment and microphone). An equal error rate less than 3% was achieved using Gaussian Mixture Models, and a slightly greater equal error rate (less than 3.5%) was achieved using Support Vector Machines with Fisher kernel and with GMM Supervector Linear kernel, against any possible acoustic conditions.