Simple Noise Robust Feature Vector Selection Method for Speaker Recognition

摘要

The effect of additive noise in a speaker recognition system is known to be a crucial problem in real life applications. In a speaker recognition system, if the test utterance is corrupted by any type of noise, the performance of the system notoriously degrades. The use of a feature vector selection to determine which speech frames are less affected by noise is the purpose in this work. The selection is implemented using the euclidean distance between the Mel features vectors. Results reflect better performance of robust speaker recognition based on selected feature vector, as opposed to unselected ones, in front of additive noise.