Testing realistic forensic speaker identification in Japanese : a likelihood ratio based approach using formants

Abstract

This thesis sets out to investigate whether or not forensic speaker identification can be performed, using forensically realistic data, which is natural and non-contemporaneous speech. To date, there is no forensic phonetic research which tests how accurately speakers can be discriminated on the basis of their voice under forensically realistic conditions, despite the fact that the use of speech recordings for forensic investigation or as a part of evidence in court is not an unusual practice today. This research thus aims to provide the first test of the accuracy of realistic forensic speaker identification using centre frequencies of formants, which are today the most commonly used acoustic parameter in actual forensic speaker identification today. The current state of forensic speaker identification in Japan also signifies this research. Forensic speaker identification in Japan has relied on visual examination of spectrograms and occasional use of the automatic speaker recognition technique. The research on forensic speaker identification also concentrates on the application of the automatic speaker recognition technique, and no linguistic analysis or interpretation of speech data has been included. This thesis therefore will serve as the first linguistic analysis in Japanese forensic speaker identification research. This thesis frrstly examines what segment f formant combinations are more promising as the speaker identification parameter. Those parameters are then incorporated, and how accurately they discriminate two speech samples is tested. For the testing, three different statistical approaches are presented and examined. As result, the distance based approach using likelihood ratio as the score for discrimination test (likelihood ratio-based distance method) was found to be most effective. The results of this testing showed that speakers can indeed be discriminated on the basis of their formant frequencies, as long as enough number of parameters are incorporated. With this approach incorporating six parameters, the successful discrimination rates were found to be approximately %.7% for positive discrimination (discriminating two different speakers) and 90% for negative discrimination (identifying the same speaker), when the threshold was set at likelihood ratio 1. Show more