Independent Component Analysis for Speech Enhancement with Missing TF Content

We address the problem of Speech Enhancement in a setting where parts of the time-frequency content of the speech signal are missing. In telephony, speech is band-limited and the goal is to reconstruct a wide-band version of the observed data. Quite differently, in Blind Source Separation scenarios, information about a source can be masked by noise or other sources. These masked components are “gaps” or missing source values to be “filled in”. We propose a framework for unitary treatment of these problems, which is based on a relatively simple “spectrum restoration” procedure. The main idea is to use Independent Component Analysis as an adaptive, data-driven, linear representation of the signal in the speech frame space, and then apply a vector-quantization-based matching procedure to reconstruct each frame. We analyze the performance of the reconstruction with objective quality measures such as log-spectral distortion and Itakura-Saito distance.