Speech variability: Minimum sample size on forensic applications

• Grant number: 24/04240-0
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: August 01, 2024
• End date: July 31, 2028
• 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

Speech is complex, dynamic, and produced in a context of variation conditioned by social and biological factors, intrinsically linguistic or external. Any theoretical framework, mathematical modeling, or acoustic measurement is subordinate to this nature. The principle supporting Phonetics and speech science considers that such variation (individual or communal) occurs within a describable limit and pattern. However, a theoretical-practical problem emerges: how to rigorously establish a speech sample -- for a given acoustic parameter -- that is representative of the population of interest, presenting the prototypical speaker's features (or community) and not a set biased by different variations, without reaching the limit assumption or capturing an existing pattern. This question and its consequences for forensic phonetics correspond to the proposed research's central point. A feasible response path consists of accumulating sufficient data and establishing the minimum sample size for variability to reach statistical stability. The methodological fundamental is generating the cumulative average of a given acoustic parameter and mathematically identifying the variance stabilization point with changepoint analysis. From a sociophonetics standpoint, we will investigate the issue -- in its technical and epistemological aspects -- with a set of 86 prosodic-acoustic descriptors with different corpora of Brazilian Portuguese, in different dialect regions, American English, Mandarin, and their synthetic equivalents, generated by artificial intelligence - given the expanding misuse of generative models with deepfakes, voice cloning.