Forensic speaker recognition in likelihood ratio framework

To test the performance of vowel cepstrum in forensic speaker recognition, a forensic speaker identification method based on likelihood ratio and Mel-frequency cepstral coefficients (MFCC) features is presented. The method is tested in vowel /a/ of 45 peoples' telephone dialog recordings and shows high identification ratio. Experimental results show that the method can identify the speaker, and quantify the evidence strength according to the acoustic difference between the questioned recording and the suspect's recording, thus providing the scientific and reasonable evaluation results to court. Compared with the manual method in formants extraction and pitch extraction, the auto extraction of features increases the efficiency and performance of forensic speaker identification system.