Forensic anthropometry from voice: An articulatory-phonetic approach

This paper addresses a problem that is of paramount importance in solving crimes wherein voice may be key evidence, or the only evidence: that of describing the perpetrator. The term Forensic anthropometry from voice refers to the deduction of the speaker's physical dimensions from voice. There are multiple studies in the literature that approach this problem in different ways, many of which depend on the availability of sufficient volumes of speech for analysis. However, in the case of many voice-based crimes, the voice evidence available may be limited. In such cases it is especially advantageous to regard the recorded signal as comprising multiple pieces of evidence. In this paper, we show how this can be done. We explain why, for any anthropometric measurement from speech, it makes sense to consider the contributions of each articulatory-phonetic unit independently of others, and to aggregate the deductions from them only in the aftermath. This approach is based on the hypothesis that the relative evidence given by different compositional units of speech can be more indicative of the anthropometric factor being deduced, than the evidence derived from the aggregate voice signal. We explain the applicability of this approach through experiments on standard speech databases.