Reliability detection by Fuzzy SVM with UBM Component feature for emotional speaker recognition

Speaker-emotion variability is one of the most significant factors inducing the degradation of the SR (Speaker Recognition) system in real life. With the change of the vocal tract shapes and the vocal fold frequencies under emotional state, some short-time acoustic features are assumed to shift from the neutral ones and turn into the unreliable segments which severely affect the performance of the SR system. Therefore, detecting and pruning those unreliable segments, i.e. reliability detection, is a way to improve the SR system. This paper is proposed to employ novel UBM Component feature to detect and prune those unreliable segments. Based on the proposed feature, we established UBM Component Fuzzy SVM (UCFSVM), which is composed of a set of Fuzzy SVM (FSVM) classifiers constructed under each UBM component. The experiment carried on MASC shows Identification Rate (IR) improvement of 3.64% compared to the baseline GMM-UBM system, and 2.82% to the traditional SVM system.