Affect-Insensitive Speaker Recognition by Feature Variety Training

摘要

A great deal of inner variabilities such as emotion and stress are largely missing from traditional speaker recognition system. The direct result is that the recognition system is easily disturbed when the enrollment and the authentication are made under different emotional state. Reynolds [1] proposed a new normalization technique called feature mapping. This technique achieved big successes in channel robust speaker verification. We extend the mapping idea to develop a feature variety training approach for affective-insensitive speaker recognition.