Short Utterance Speaker Recognition Using Time-Delay Neural Network

Abstract

In recent years, important studies on speaker recognition have been implemented. Some solutions have been proposed and high achievements have been achieved. But one of the major issues faced by speaker recognition researchers is the short utterance speaker recognition. In short utterances, the recognition performance decreases. Within the scope of this study, it is aimed to increase the speaker recognition performance in short-term utterances by using Time-Delay Neural Networks (TDNN). I-vector-based systems have been developed using conventional GMM-UBM and TDNN-UBM-based methods. In this study, error rate changes of the audio files at various durations are compared using GMM and TDNN methods. Our results demonstrate that TDNN method provides improvement in equal error rate(EER) values compared to GMM method on speaker recognition area. However, the relative improvement of TDNN decreases while the test data duration decreases.