Formant trajectories as indices of phonetic variation for speaker identification

Abstract

Formant trajectories are known to acoustically encode static and dynamic aspects of speech, which have potential to identify the speaker, such as differences in vocal tract morphology, articulatory setting, dialect, and speaking style. Although widely used in studies of socio-phonetic variation, formant trajectories have not been extensively investigated for purposes of voice identification. An exploratory investigation was conducted to assess the speaker discriminating potential of formant trajectory marching, in a group of fifteen male speakers tested one month apart, using phonologically matched sonorant segments of speech tagged for the presence of connected speech processes (CSPs) and level of linguistic stress. Identification rates of 85 to 93 per cent were achieved on the basis of the nine best discriminating segments (2 to 3 seconds of speech). Speaker differences in the application of CSPs did nor appear to contribute significantly to the speaker discrimination power of the formant trajectories. Stress was relevant, in terms of how well segments were processed by the formant tracker; a finding which was supported by regression analysis of segment durations and amplitudes on speaker discrimination scores. Implications for forensic speaker identification are discussed.