Parameters for speaker comparison: automatic extraction and analysis for forensic purposes

• Grant number: 19/24234-7
• Support Opportunities: Scholarships in Brazil - Master
• Start date: January 01, 2021
• End date: May 31, 2022
• Field of knowledge: Linguistics, Literature and Arts - Linguistics
• Principal Investigator: Plinio Almeida Barbosa
• Grantee: Niasche Moraes de Aquino Silva
• Host Institution: Instituto de Estudos da Linguagem (IEL). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

The earliest cases in which speech was used as forensic evidence date back to the 17th century. Years after with a methodological-technical refinement, the speech analysis of different suspects was called Speaker Comparison (SC). The biggest challenge of SC is that unlike DNA and fingerprints, there is no specific speech characteristic which is biologically determined, unalterable, and invariant (Foulkes & French, 2012). To solve the problem and guarantee an acceptable degree of confidence in the analysis, several parameters are used for SC. Despite the gain in analytical robustness, given the complexity and time cost of correlating different parameters, two inherent needs in forensic phonetic arise: to investigate/formulate possible new parameters and to expand automatic extraction/analysis techniques. Both constitute the central theme of this project which is divided into three stages: (I) evaluate and select acoustic parameters pertinent to SC, (II) produce scripts for automatic parameter extraction and graphical plotting and (III) test the extraction quality in several signal-to-noise relationships and a practical implementation of the script in two forensic case simulations. The first stage corresponds to an extensive literature review and pilot tests using the corpus, C-ORAL-BRAZIL (Raso & Mello, 2011), given several criteria for selection and formulation of optimal parameters for SC (Rose, 2002). Then, from the parametric description, we will produce scripts using the Praat Scripting Language (Boersma & Wenick, 2009). In the last step, we will add different noises to the corpus data and statistically analyze the behavior of the script. The first simulation will consist of a blind test. From a set of 10 corpus speakers, one will be selected as a "criminal" to be identified by SC with the developed script. The second simulation is a comparison between original and synthesized, by artificial intelligence systems, audios. (AU)