Odysseus

Forensic speaker identification involves the assignment of a probability that a sample of speech (for example a recorded phone call) came from a specified person from whom other speech samples (though not necessarily text-identical) may be obtained. This is still an unsolved task, on which many researchers are working. We have collected a unique speech database in which multiple speakers are recorded in a variety of conditions. Along with control recordings of normal read speech, speakers read a variety of texts along with other speakers, which forces each speaker to approximate many characteristics of their co-speaker's voice. We can use these recordings to look for hallmarks of a given speaker which do not change, evan as the speaking situation changes considerably.

In this project, the student will be introduced to the speech corpus, and provided with a selection of tools for feature extraction. There is a deal of trial and error in assessing the utility of various feature sets for picking out specific speakers from the set. The project will involve trying out a broad range of features, and assessing their stand-alone and combined utility at identifying speakers from a set of about 40. The student will learn to adopt a principled methodology in searching a large space of possible features, and will apply consistent test methods in evaluating the utility of each feature tested.

Relevance of Project to the Host Laboratories:

This project complements ongoing work in the group aimed at identifying features which can serve for text-independent speaker identification under noisy channel conditions. The student's investigations may suggest novel features for exploration, or, more likely, they will help us to identify almost optimal sets of features from a known set.

Supervisors:

Dr. Fred Cummins (Computer Science and Informatics, UCD)

Keywords:

Speaker Identification, Speech Analysis