BROAD PHONETIC CLASSES EXPRESSING SPEAKER INDIVIDUALITY
Margit Antal, Gavril Toderean · 2006
Vector quantisation and Gaussian mixture modelling methods are very popular methods for automatic speaker identiflcation. First we give a concise overview of these methods, then present some measurements com- paring them on behalf of the TIMIT corpus. The aim of this paper is to study the in∞uence of the speech material on performances of such methods. For this purpose pure phonetic speaker models were created containing speech data from a single broad phonetic class. The speaker discriminative prop- erty of these pure phonetic speaker models had been investigated. Among the broad phonetic classes nasals and vowels were found to be particularly speaker speciflc.