Non-linear Power Exponent Effect in GFCC for Bangla and Malay speech Separation

The applications of automatic speech processing are increasing due to the increasing demand for human machine interaction services in our daily activities. The human-machine interactivity is hampered under noisy conditions. Nonlinearities in the human auditory system make it noise-robust. Inspired by the performance of the auditory system, conventional auditory front-ends apply either power exponent or log operation to emulate cochlear compression or suppression effect. For example, the Gammatone Frequency Cepstral Coefficient (GFCC) applies the cube root power exponent to add a cochlear suppression effect in it. In this work, I investigate the effect of power exponent nonlinearity on the performance of the GFCC. I evaluate the performance through the separability of Bangla and Malay accent-based speeches. The simulated result shows that the choice of power exponent nonlinearity significantly contributes to producing a noise-robust performance.