Speaker recognition on mono-channel telephony recordings

Conversations stored as mono data is a common problem inmany real world speaker recognition applications. In this paper,we focus on investigative scenarios, where a number ofmono telephone conversations are available for a speaker of interest.For example, a human operator may have verified thatthe speaker is present in these conversations. We propose severalapproaches for automatically creating enrollment modelsfor the speaker of interest from such data. We then use the enrollmentmodels to search for appearances of the speaker of interestin other calls. We analyze the performance of the differentmethod on two dataset that matches our scenario, one is from asimulated case and one is from a real case.