Speaker identification using discriminative feature selection: a growing neural gas approach

A new method of text-dependent speaker identification using discriminative feature selection is proposed. The characteristics of the proposed method are as follows: feature parameter extraction, vector quantization with the growing neural gas algorithm, model building using Gaussian distributions and discriminative feature selection according to the uniqueness of personal features. The speaker identification algorithm is evaluated on a database that includes 25 speakers each of them recorded in 24 different sessions. All speakers spoke the same phrase for 240 times. The test results showed that both the false rejection rate and false acceptance rate were under 1%. The overall performance of the system was 99.5%.