Title: Compensation of variability using median and i-vector + PLDA for speaker identification of whispering sound
Authors: Vijay M. Sardar
Addresses: Department of Electronics and Telecommunication, Jayawantrao Sawant College of Engineering, Pune, Maharashtra, India
Abstract: Speaker identification from the whispered voice is troublesome assignment contrasted with neutral as the voiced phonations are missing in the whisper. The success of the speaker identification system mainly depends upon the selection of appropriate audio features. The various available audio features are explored here and reported that the timbre features outperform to identify the whispering speaker. Only the well-performing, and thus limited timbre features are sorted by Hybrid Selection Algorithm. The timbre features named Brightness, Roughness, Roll-off, MFCC and irregularity using CHAIN database offer improvement in the identification outcomes by 5.8% over the baseline system. The framework ought to be robust enough to repay intra-speaker and inter-speaker variability, including channel impacts. The analysis using timbre features based on Median value predicted that the intra-speaker variability is being compensated. The use of median timbre features reported further enhancement of 1.12% compared to using timbre features and a further decline in False Negative Rate (FNR). The use of i-Vector + Probabilistic Discriminant Analysis (PLDA)-Support Vector Machine (SVM - cosine kernel) contributed the relative improvement in accuracy by 8.13%. The reduction in False Positive Rate (FPR) and False Negative Rate (FNR) confirms better variability compensation.
Keywords: whispered speaker; median timbre feature; i-vector; cosine kernel; support vector machine.
International Journal of Computer Applications in Technology, 2021 Vol.67 No.1, pp.47 - 57
Received: 11 Apr 2020
Accepted: 08 Dec 2020
Published online: 07 Feb 2022 *