Un metodo statistico per il riconoscimento del parlatore basato sull'analisi dei formanti


  • Tommaso Bove Servizio Polizia Scientifica, Roma
  • Paolo Emilio Giua Istituto di Acustica O. M. Corbino - CNR Roma
  • Alessandra Forte Università degli Studi di Roma "Tor Vergata"
  • Carla Rossi Università degli Studi di Roma "Tor Vergata"




In this paper, a method for the forensic identification of speakers is presented, based on the analysis of the pitch and the first three formants of the four vowel: "a", "e", "i" and "o". Using these data, the method estimates the probability density function (pdf) of the Mahalanobis distance both of the defendant from himself (intra-distance estimation) and from the voices of the control set (inter-distance estimation), using the Kernel method for each vowel. The Mahalanobis norm is then used to estimate the pdf related to the four vowel. The sample under study is then classified according to the Maximum Likelihood Criterion approach. This allows one to estimate a unique decision threshold and the probabilities of the two possible classification errors (false acceptance and false rejection). The method has been applied to real data provided by Police Scientific Service of Rome, in the framework of a European Project.




How to Cite

Bove, T., Giua, P. E., Forte, A., & Rossi, C. (2002). Un metodo statistico per il riconoscimento del parlatore basato sull’analisi dei formanti. Statistica, 62(3), 475–490. https://doi.org/10.6092/issn.1973-2201/420