Intelligibility Improvement using Kalman Filtering em Approach in Formant Shifting Framework

Abstract

A formant shifting (FS) based speech intelligibility improvement approach [1] performs modifications on clean vocal sound messages. However in real time conditions, clean speech signal is generally mixed with car noise environment and reverberations. The LP Coefficients (LPCs) of clean speech signal is generally unavailable. Hence in this work, we would like to estimate LPCs from the mixed noisy signal using a combination of Kalman Filtering and Expectation Maximisation (KF-EM) based recursive algorithm. Once LPCs of clean speech from mixed noisy signal are estimated, formants are then computed and shifted away from the region of competing noise. In comparison to traditional LP coding analysis (LPCA) [1] and repeated autocorrelation (RA) [2] based formant estimation methods, the proposed KF-EM algorithm accurately estimates the LPCs from the noisy signal and subsequently, accurate formants are estimated. This results in significant intelligibility improvement in car noise environment for proposed algorithm, when used in FS framework. The subjective and objective analysis on NOIZEUS database are used to evaluate the performance of the proposed method. © 2019 IEEE.