Consistent speaker identification via Wigner smoothing techniques

A voiceprint database for automatic speaker identification is proposed. With traditional short-time Fourier transform techniques system limitations arise not from discrimination between two different speakers but from single-speaker variations. The proposed estimator uses a derivative of the Wigner distribution (WD) function called the smoothed-discrete-WD (SDWD). The SDWD is shown to produce a significantly more consistent estimate over the differing samples of a given word utterance when the DWD kernel has been averaged over the entire window. The formant frequencies are well identified and remain consistent over the samples.<>