Text and language-independent speaker recognition using suprasegmental features and support vector machines

In this paper, presence of the speaker-specific suprasegmental information in the Linear Prediction (LP) residual signal is demonstrated. The LP residual signal is obtained after removing the predictable part of the speech signal. This information, if added to existing speaker recognition systems based on segmental and subsegmental features, can result in better performing combined system. The speaker-specific suprasegmental information can not only be perceived by listening to the residual, but can also be seen in the form of excitation peaks in the residual waveform. However, the challenge lies in capturing this information from the residual signal. Higher order correlations among samples of the residual are not known to be captured using standard signal processing and statistical techniques. The Hilbert envelope of residual is shown to further enhance the excitation peaks present in the residual signal. A speaker-specific pattern is also observed in the autocorrelation sequence of the Hilbert envelope, and further in the statistics of this autocorrelation sequence. This indicates the presence of the speaker-specific suprasegmental information in the residual signal. In this work, no distinction between voiced and unvoiced sounds is done for extracting these features. Support Vector Machine (SVM) is used to classify the patterns in the variance of the autocorrelation sequence for the speaker recognition task. © 2009 Springer Berlin Heidelberg.