Abstract:
A new algorithm for speaker identification is suggested. A set of sound files belonging to two speakers is given. There are two short files known to be corresponding to two different speakers. The speakers are not supposed to have articulated the same phrase. The task is to establish belonging of each file in the set. The problem is solved using an approximation of spectral surface of sound file by wavelets of special kind. The compressed form of the spectral surface is processed by a neuron net. The decision about the belonging is made basing on the values produced by the neuron net. Some results of an experiment with files from a speech database are presented.
Keywords:speaker distinguishing, short phrases, graphical method.
Fulltext:
PDF file (263 kB)