Visual processing-inspired Fern-Audio features for Noise-Robust Speaker Verification

摘要

In this paper, we consider the problem of speaker verification as a two-class object detection problem in computer vision, where the object instances are 1-D short-time spectral vectors obtained from the speech signal. More precisely, we investigate the general problem of speaker verification in the presence of additive white Gaussian noise, which we consider as analogous to visual object detection under varying illumination conditions. Inspired by their recent success in illumination-robust object detection, we apply a certain class of binary-valued pixel-pair based features called Ferns for noise-robust speaker verification. Intensive experiments on a benchmark database according to a standard evaluation protocol have shown the advantage of the proposed features in the presence of moderate to extremely high amounts of additive noise.