A new feature extraction based the reliability of speech in speaker recognition

The paper discusses the reliability of speech feature extraction and its application to speaker recognition. Usually, a speaker recognition system consists of a front-end feature extractor and a back-end classifier. The usual speech feature extractor only extracts the feature parameters, it does not estimate the reliability of these parameters. We propose a new speaker recognition method based on the reliability of speech features extracted. We apply a different weight to each feature vector according to the estimated reliability of this vector and then determine its role in speaker recognition. Our experiments clearly show that this strategy improves the effectiveness in text-independent speaker identification.