Authors : Jhon Kevin Segura, Diego Renza and L. Dora M. Ballesteros
Abstract: In literature, there are several manuscripts related to finding the best feature or the best classifier for audio verification systems. However, cross validation with both criteria has not been widely discussed. In this research, 15 classifiers and six features have been selected to obtain ninety options for audio forensics verification. The aim is to provide suggested combinations for forensics researches. The evaluated classifiers are based on decision trees, discriminant analysis, support vector machines, nearest neighbour and hybrid classifiers. The feature extraction is based on Mel-Frequency Cepstral Coefficients (MFCC) and cochleagrams, using principal component analysis optionally. The tests are performed on a database of 50 speakers and 10 utterances per speaker and the assessment of classifiers is made by means of accuracy. According to the results, the best combination is MFCC with linear discrimination, followed very close by MFCC+PCA with linear discriminant.
Jhon Kevin Segura, Diego Renza and L. Dora M. Ballesteros, 2018. Cross Validation of Machine Learning Classifiers and Features for Audio Forensics Verification. Journal of Engineering and Applied Sciences, 13: 4512-4517.