ESCA Workshop on Automatic Speaker Recognition, Identification, and Verification
In this paper we report about a set of experiments in speaker verification, based on algebraic methodologies that consider the covariance matrix of a given speech parametrisation, like LPC cepstrum, vocal tract area and so forth, as the reference "holder" of the information related to the speaker's identity. We use an Italian 12 speakers telephonic database, recorded many years ago (first and second collation within two years, third collation after ten years). Speakers are all of the same regional area (Padova) and with very similar voices. We test three different methods on various speech parametrisations and for several selections and grouping strategy of speech file. We found that integral scoring methods as the Arithmetic-Harmonic Sphericity Measure give tantalising results, that standard Malahanobius-like distance has high diagnostic properties, while other Karhunen-Loeve based methods, that theoretically has a high possibility, give disappointing results. The results of these experiments give us guidelines to more extensive investigation on wider databases, focusing our attention to an analytical study of the different properties of algebraic methods in speaker verification.
Bibliographic reference. Falcone, Mauro / Paoloni, Andrea (1994): "Text-independent speaker verification based on multiple reconstruction of selected speech zones", In ASRIV-1994, 173-176.