Speaker recognition using Kohonen's self-organizing feature map algorithm

We used the Kohonen self-organizing feature mapping algorithm to derive speech templates for text-independent automatic speaker recognition. The speaker recognition algorithm is based on template matching. The Kohonen method of deriving templates was compared with an alternate method based on cluster averaging. We found the recognition performance of the two methods to be about the same, given equal computation. However, the Kohonen method has a practical advantage that the desired number of templates is specified in advance. This advantage can be significant, particularly for noisy or distorted speech data, because it avoids the need for the operator to 'tune' the system for the input data.