LAUSR

Recent work has shown that a machine learning algorithm can produce large speech intelligibility in noise increases for hearing-impaired listeners. This algorithm involves a deep neural network trained through supervised learning to estimate the ideal binary or ratio mask. The direct translational potential of this work is addressed currently. Primary issues surrounding future implementation into hearing aids and cochlear implants involve (i) the ability to generalize to conditions not encountered during training, and (ii) the computational load associated with operation of such an algorithm. Substantial advances have been made with regard to generalization. These will be outlined as will associated decisions that can be made. The computational load associated with training and operation of a network will also be addressed. Here, we propose an alternative implementation that offers multiple advantages over current approaches.