Performance evaluation of speaker recognition system using area under ROC curve for extracted novel features from SDM and MDM speech signals
Sukhvinder Kaur, Chander Prabha · AIP conference proceedings · 2022
An important field of artificial intelligence is speech processing that includes speech recognition, speaker recognition, and speaker diarization. One of the important aspects of human-machine speech interface technology is speaker recognition (SR) which is used in a biometric system. It automatically recognizes the speaker, based on speaker- specific information included in the speech. The main steps of SR are feature extraction, feature matching, and performance evaluation. The database required for its processing is acquired from a personal digital assistant (PDA) that includes two ways to record speech signals: Multiple Distant Microphone system (MDMS), and Single Distant Microphone system (SDMS). In this research work, a novel feature extraction algorithm based on “Daubechies wavelet” and “Teager Kaiser Energy operator (TKEO)” has been applied on both SDMS and MDMS speech signals. Next, feature matching of different features has been reckoned with a distance metric algorithm using the Bayesian Information Criterion (BIC). At the end, the enactment of the SR system is estimated by applying “Receiver operating characteristics (ROC)”, area under curve, and “Equal Error Rate”. It has been concluded that the proposed algorithm with MDMS claims 85.17 % of accuracy and 96.28% of area under curve which is better than that of SDMS performance.