This paper presents theoretical aspects and experimental results relating to the use of second-order statistical measures, for text-independent speaker recognition. These measures are easy to implement, and computationally efficient. They provide remarkable results for the 420 speakers of the TIMIT database.