Testing the validity and reliability of forensic voice comparison based on reassigned time-frequency representations of Chinese /iau/

摘要

Time-frequency reassigned representations of vowel segments have repeatedly been proposed as features for performing forensic voice comparison (FVC). Combined with a thresholding on the second-order mixed partial derivative of the short-time Fourier transform (STFT) phase to reduce artifacts and noise (pruning) such representations offer a sharpened representation of AM/FM components in the speech signal. We test the validity and reliability of FVC systems based on pruned reassigned time-frequency representations extracted from manually labeled /iau/ segments from a database of voice recordings of 60 female Chinese speakers. Logistic-regression fusion was used to combine the systems with a baseline MFCC GMM-UBM system. Performance was assessed as degree of improvement over the baseline system. Three recording channel conditions were tested: high-quality v high-quality, mobile-to-landline v mobile-to-landline, and mobile-to-landline v high-quality. For the latter two conditions improvements over the baseline system were observed in terms of validity, but with decreases in reliability.