Abstract—In the last time, the forensic speaker recognition has focused the attention and interest of the scientific community, the voice signal present a significant challenger due to its large variability, but in the forensic science, the variability is even more complex due to multiple factors that are present in this area of forensic biometrics, e.g., short time recordings, noise environmental, channel mismatched, non-contemporaneous recordings and so on. In this paper we study the performance of the Mel Frequency Cepstral Coefficients, using Multitaper Spectrum Estimate with promising preliminary results. This technique reduces the variance and improves the performance in forensic applications. In this work we focus in non-contemporaneous recordings, to modelling we use Gaussian Mixture Models.
Index Terms—Forensic speaker recognition, GMM, MFCC, multitaper analysis.
The authors are with Departamento de Procesamiento de Señales, Facultad de Ingeniería, UNAM(e-mail: jtrangol@yahoo.com, abelhc@hotmail.com)
[PDF]
Cite:Jose Trangol and Abel Herrera, "Low Variance Non-Contemporaneous Recordings Using Multitaper Analysis in Forensic Applications," International Journal of Modeling and Optimization vol. 3, no. 5, pp. 425-428, 2013.