Topological voiceprints for speaker identification

Despite its noninvasive nature, subject identification by voice is not as popular as other biometric procedures (i.e. fingerprinting). In part, this is due to the difficulty of establishing how close is close enough when comparing spectral features. In this work, we address this issue by showing how to characterize spectra by means of sets of integers, borrowing topological tools used in the theory of dynamical systems. On the other hand, we report an empirical result: within a relatively small bank of speakers, there are subsets of integers that seem to strenghten the speakers' identity information. These results suggest a new direction in the identification of subjects by voice: one in which arrangements of integers define voiceprints that stand on their own, despite any acceptance/rejection thresholds.