Frequent-words analysis for forensic speaker comparison

Abstract

While many authorship analysis techniques could be applied to any type of linguistic data, there is very limited research on their application on transcriptions of spontaneous speech to help identify the speaker. In authorship analyses, frequency analysis of function words has been quite a successful feature since, when writing a text, people tend to use different function words in different frequencies. Thus, this feature is sensitive to author style while it remains relatively insensitive to the topic of the text. In this paper, we make the cross-over to speech samples and test whether frequency analysis of the most common words in speech, which are mostly function words, has potential as a speaker discriminant. We propose a method within the likelihood ratio framework for court applicability. Our approach takes into account both the similarity and typicality of word frequencies by employing a method using percentile-rank for feature extraction. We apply our method on the forensically relevant dataset FRIDA, achieving good results even when a limited amount of data is available.