Nasal spectra for forensic voice comparison

For features to be effective in forensic voice comparison, they must have relatively low within-speaker variability and relatively high between-speaker variability. An understudied source of features, which potentially meets these criteria is the acoustic spectrum of nasals. Nasals spectra contain poles and zeros dependent upon nasal cavities. The latter are complex static structures which vary from person to person. Theoretically, nasal spectra may therefore have low within-speaker and high between-speaker variabilities. This study evaluates different methods for extracting spectral features (e.g., pole-zero models, all-pole models, and cepstra) and using them as part of a likelihood-ratio forensic-voice-comparison system. The validity and reliability of each system is empirically evaluated using /m/ and /n/ token extracted from a database of voice recordings of 60 female speakers of Chinese.