A multi-step method for speaker identification

In the context of the countermeasures against criminal or terrorist acts, the attribution of identity to a unknown speaker, (for example to an individual talking on a phone line), may play a primary role. Speaker identification (SI) may be performed with or without the human support and, according to this distinction, SI systems are divided in "semi-automatic" and "automatic" (J. P. Campbell, Sept. 1997). In semi-automatic protocols, the process of identification is carried out by means of electronic instruments with the support of a technician who generally has a linguistic background. Automatic systems do not need human support and may operate in quasi-real-time, and this may represent a feature particularly appealing in some operative scenarios. Obviously, the complexity of automatic systems is relevant and then, generally, complex architectures are required. In the present paper the authors propose a four-classifiers methodology which exhibits some innovative solutions in the context of similar approaches. In particular, a new robust approach to pitch extraction allows to overcome a set of problems generally associated with this task