Forensic voice comparison using sub-band cepstral distances as features: A first attempt with vowels from 306 Japanese speakers under channel mismatch conditions

This study presents the latter part of an exploratory study of the potential of sub-band parametric cepstral distance (PCD) as an alternative forensic voice comparison (FVC) feature to formants and cepstral coefficients. Using 5 Japanese vowels produced by 306 male Japanese speakers, we conducted LRbased FVC experiments under a channel mismatch condition, with sub-bands selected in reference to the expected formant locations. Combining 3 sub-band PCDs from F1, F2, and F3 ranges, sub-band PCDs outperformed the full-band PCDs in speaker classification, demonstrating their promise as an automatically extractable, robust, and linguistically interpretable acoustic feature for FVC. Index Terms: Sub-band cepstral distance, likelihood ratio, forensic voice comparison, channel mismatch, Japanese vowels