Use of Gaussian Mixture Models in Macedonian forensic speaker identification

Forensic Speaker Identification is one of the methods used in court to identify the speaker in an incriminating recording from a set of verifiable speakers. It uses the characteristics of the speaker´s voice to differentiate the speakers, such as formant frequencies and bandwidths, pitch, segmental timings etc. A methodology is proposed comprising formant frequency feature extraction, Gaussian Mixture Model (GMM) modeling, and various similarity measures, for the task of Macedonian forensic speaker identification. A novel use of the scalar product (SP) of functions as a similarity measure of two GMMs is also proposed. A test-bed database is created comprising a set of recordings from 18 suspects and one incriminating recording to test the proposed methodology. The results have shown reliable forensic speaker identification, and are better than the professional forensic software that was available for comparison.