A first attempt at compensating for effects due to recording-condition mismatch in formant-trajectory-based forensic voice comparison

This paper reports the results of a first attempt at compensating for variability in formant-trajectory representations due to differences in suspect and offender recording conditions. Formant-trajectory measurements were made on tokens of /iau/ in a database of high-quality and mobile-to-landline-transmitted recordings of 60 female speakers of Chinese. Discrete cosine transforms (DCT) were fitted to the formant trajectories. DCTs were transformed using feature mapping and using canonical linear discriminant functions. Transformed coefficients were used to calculate likelihood ratios via the multivariate kernel density formula. Application of the compensation procedures did not lead to substantial improvement in validity or reliability.