Time-Frequency Feature Extraction from Residues of Voiced Phonemes

A speaker -recognition system attempts to recognize a speaker by his/her voice. A common problem in these systems is that a mismatch in the training and testing conditions damages the performance a great deal. The objective of this paper is to seek for new features of speech that can help to develop alternative techniques to identify who speaks. Being speech the result of a dynamic process and trying to capture its no stationary nature, the temporal evolution of the vocal tract parameters is analyzed and then a joint time-frequency analysis of the excitation source of isolated vowels from different speakers is made. Considering the hypothesis that important speaker information is in the glottal flow, Flanagan and Fant source-filter model for voiced phonemes is taken as a starting point where the residues are an acceptable representation of the excitation signals. Then they are analyzed with joint time -frequency techniques. Taking into account the analogy between a time- frequency representation and bidimensional probability densities, and that these functions can often be described by some of their moments, it is expected that the joint time-frequency density can be described by its low order moments. Even if results are at the experimental stage, they support this hypothesis. Although it is not expected that joint time-frequency analysis can completely solve the problem, the incorporation in the feature vector of some components of this type, will certainly improve the systems currently in use.