Open-Set Speaker Identification pipeline in live criminal investigations

Abstract

Speaker recognition has many applications in conversational data, including in forensic science where Law Enforcement Agencies (LEAs) aim to assess the identity of a speaker on a specific recorded telephone call. However, speaker identification (SID) systems require initial enrollment data, whereas LEAs might start a case with text or video evidence, and few to no enrollment data. In this paper, we introduce the ROXANNE simulated dataset, a multilingual corpus of acted telephone calls following a screenplay prepared by LEAs. We also present a process to build criminal networks from SID, by addressing practical constraints of these investigations. Our process reaches a speaker accuracy of 92.4% on the simulated data and a conversation accuracy of 84.9%. We finally offer some future directions for this work.