Hybrid solution to single-channel hybrid noisy speech for an industrial environment

The aim of this research is developing a dynamic automatic noisy speech recognition system (DANSR) to recognize small spoken commands in a hybrid noisy industrial environment. For this we first focus on the noise problem while most noisy speech recognition studies focus on enhancing the noisy speech features. By hybrid noise we understand the environmental mixed noise which is generated from different sources. We discriminate the noise as strong, time-varying steady-unsteady, mild. The hybrid noise has different loudness varying from extremity to mild and it affects the delivered spoken commands at varying extent during its lasting time. The hybrid solution is a combined innovative approach to a long existing problem. Here we have only one input that is mixed and we expect its single output. We have only one microphone and therefore we are working on a single-channel only. We treat strong noise as outliers and for this we present an innovative treatment. We employ Kalman filter (KF) for time-varying steady-unsteady noise problem in the M-band signal. The signal is based on a fast modified covariance method of linear prediction as an unconstrained least squares (MULS) approach. The time-varying steady-unsteady noise is modeled by Yule-Walker approach and updated in each band. Finally we have a principle component analysis (PCA) as a solution to the mild noise.