Teager Energy Cepstral Coefficients for Classification of Normal vs. Whisper Speech
Kuldeep Khoria, Madhu R. Kamble, Hemant A. Patil · 2020
The whispered speech is quite different from natural speech in the context of nature, acoustic characteristics, and generation mechanism. In order to improve the robustness of Automatic Speech Recognition (ASR) system, it is very important to analyze the mismatched training and testing situations and propose a robust acoustic features to enhance the whisper recognition. In this paper we propose to use Teager Energy Cepstral Coefficients (TECC) which uses Teager Energy Operator (TEO) for estimating "true" total energy of the signal, i.e., the sum of kinetic and potential energies which is contradictory to the traditional signal energy approximation, which only takes kinetic energy into account, i.e., L2norm of the signal. In this study, experiments are performed on wTIMIT and CHAINS corpus. For wTIMIT corpus, frame-level accuracy of 92.22 % is obtained and for CHAINS corpus, it is 95.61 %. We have also estimated the performance measure of the classifier by using Matthew Correlation Coefficient (MCC), F-measure, and J-statistics. Furthermore, experiments are performed by considering latency period from a practical deployment viewpoint, and the trade-off between latency period vs. accuracy is discussed for both the corpora.