Analysis of the Effect of Speech-Laugh on Speaker Recognition System

Abstract

A robust speaker recognition system should be able to recognize a speaker despite all the possible variations in speaker's speech. A common variation of the neutral speech is speechlaugh, which occurs when a person is speaking and laughing, simultaneously. In this paper, we show that speech-laugh significantly degrades the performance of an i-vector based speaker recognition system. Further, we show that laughter and neutral speech contain complementary speaker information, which can be combined to improve the performance of the speaker recognition system for speech-laugh scenarios. Using AMI meeting corpus database, we show that by including neutral speech and laughter in enrollment phase, the performance of the system in the speech-laugh scenarios can be relatively improved by 36% in EER.