Provided by:International Journal of Scientific Research Engineering &Technology (IJSRET)
Presence of additive noise and room reverberations combined together offer challenge to build robust speaker identification system to suit real world applications. This paper tries to solve this issue in two steps. The speech signal corrupted by noise and room reverberations was preprocessed through binary masking using a deep neural network classifier in the first step. The room reverberations are proposed to suppress by fusing bounded marginalization and direct masking in the next step. The proposed scheme shows better speaker identification performance for various signal-to-noise ratios and reverberation times.