 |
7th Annual IEEE Information Assurance Workshop 21-23 June 2006 "The West Point Workshop" United States Military Academy, West Point, New York |
| Chair: LTC Ronald Dodge, Ronald.Dodge@usma.edu | |
|
|
|
A multi-step method for speaker identificationMario Savastano, Angelo Luciano, Dora Pagano, Bruno Peticone and Luisa RiccardiThe 7th IEEE Information Assurance Workshop (IAWorkshop 2006)West Point, New York, USA, June 21-23, 2006
AbstractIn this work the Authors present an innovative approach to Speaker Identification (SI). In particular, the methodology proposed belongs to the category of the closed-set automatic SI systems, which attempts to recognize an unknown speaker from a set of known speakers. Automatic SI has traditionally been considered particularly demanding and, nowadays, it is commonly assumed that only an appropriate combination of more classifier systems may improve the robustness of the identification. According to this approach, the Authors propose a four-step serial architecture based on an innovative set of classifiers. Three blocks (detecting pitch, power and maximum autocorrelation value) have a “whitening” purpose and restrict the number of candidates while the fourth block (assessing the frequency of the signal) is the real classifier of the system. At the end of the process, by means of a robust calculation of the distance between the unknown speaker and the “whitened” set of speakers, the system attributes a score to the candidates and, consequently, indicates the probable identity of the unknown speaker. Since the tests have been intentionally carried out on a custom database and not on one of the several corpora available on the market, at the present moment, it is difficult to provide absolute numbers in terms of accuracy of the system. In any case, the first experimental results, acquired in operatve scenarios, are extremely comforting and confirm the robustness of the system’s architecture and structure.
|
|
