In this paper we analyze speech for low-level cognitive features using linear component analysis. We demonstrate generalizable component 'fingerprints' stemming from both phonemes and speaker. Phonemes are fingerprints found at the basic analysis window time scale (20 msec), while speaker 'voiceprints' are found at time scales around 1000 msec. The analysis is based on homomorphic filtering features and energy based sparsification.
|Publication status||Published - 2005|