Statistical Model of Acoustic Parameters for Speaker Recognition

摘要

A set of 28 parameters was studied, including the first 4 formant frequencies of each Italian vowel, the fundamental frequency of each vowel, and the time durations of the plosives k, p and t. An experiment was carried out selecting (four) male individuals who recorded for 20 repeated utterances in a period of 6 months. From the recorded material a multivariate statistical analysis was carried out, determining the mean vector (centroid) for each individual and a variance-covariance matrix. The reduced models, using only 12 or 17 selected variates, are found to be more effective while the discrimination ability is not significantly impaired. The first and third formant frequencies and the plosive k duration are shown to be the most efficient parameters.