Fig 1. The DNN-based system.

Noisy speech was decomposed by a gammatone filterbank and AMS features were extracted per subband. The AMS features were fed into an DNN with two hidden layers of 128 nodes each. The system estimated a time-frequency mask (either an IBM or an IRM), and the mask was subsequently applied to the subband signals of the noisy speech, as illustrated by the dashed line, in order to reconstruct the speech signal.