Addressing database mismatch in forensic speaker recognition with Ahumada III: a public real-casework database in Spanish

摘要

This paper presents and describes Ahumada III, a speech database in Spanish collected from real forensic cases. In its current release, the database presents 61 male speakers recorded using the systems and procedures followed by Spanish Guardia Civil police force. The paper also explores the usefulness of such a corpus for facing the important problem of database mis-match in speaker recognition, understood as the difference be-tween the database used for tuning a speaker recognition sys-tem and the data which the system will handle in operational conditions. This problem is typical in forensics, where variabil-ity in speech conditions may be extreme and difficult to model. Therefore, this work also presents a study evaluating the im-pact of such problem, for which a corpus quoted as NIST4M MultiMic MisMatch) has been constructed from NIST SRE 2006 data. NIST4M presents microphone data both in the enrolled models and in the test segments, allowing the genera-tion of trials in a variety of strongly mismatching conditions. Database mismatch is simulated by eliminating some micro-phone channels of interest from the background data, and com-puting scores with speech from such microphones in unknown testing conditions as usually happens in forensic speaker recog-nition. Finally, we show how the incorporation of Ahumada III as background data is useful to face database mismatch in real-world forensic conditions.