Abstract:

The need to compare non-contemporary speech samples is a common issue in forensic speaker recognition. Even if the size of the delay is known a crucial question has still to be
answered in every single case: Is it possible that physiological aging or other factors have caused alterations of parameters of voice, speech and language to such a degree that the material basis for identification by human listeners and/or by automatic systems may no longer be regarded as adequate when a new identification task is due? Some studies on auditory speaker recognition suggest that, other variables being equal, a time lag of up to 6 years does not seem to pose a problem. To this author’s knowledge, the question has not yet been directed towards automatic identification systems. The present study uses speech
data from ten male speakers recorded at intervals of 11 years and analyses the effect of the delay in terms of (a) the ability of different groups of listeners to detect it, and (b) its
influence on the performance of an advanced automatic speaker identification system for forensic applications. Using contemporary samples for both ‘known’ and ‘unknown’
models as a benchmark the automatic speaker identification system identified all speakers correctly with LRs between 102 and over 108. In the non-contemporary condition,
where the older samples (i.e. ‘younger’ speakers) were used for the construction of models for ‘known’ and the samples recorded 11 years later for ‘unknown’ speakers, LRs of nine
speakers remained unchanged or dropped only slightly. The LR for one speaker dropped sharply. It seems that this is the only case in which vocal aging, but perhaps also other
time-related factors, may have played a rôle. The main conclusions of these experiments are that for most male speakers a delay of the size of a decade between voice samples will not pose a problem to either auditory or machine-based speaker identification.

Author: Hermann J. Künzel