3.3 Method of detection
LFCC statistical matirx as well as formant matrix is used in this paper as feature, and SVM classification is adopted, dececting system for tone disguising is shown on pic(4)
Two processes is included in SVM classification: training and testing. During training, the data contains two sets of voice data, one is the original voices, the other is.disguised voices跟压，which constitute 10 two class classifiers(change factors differs from 4 to 8, -4 to -8), each of the classifier joins the testing process. During testing, tone features is collected and tested by the 10 classifiers, Unless all of the classifiers indicates the tested tone is original ,then conclusion could be made that this tone wasn't disguised, Or else it is disguised.
4.1 Experimental environment
TIMIT is used in this paper as source, which includes 6300 pieces of voices recorded by 630 persons. The File layout is wav, sample frequency is 8000Hz, and is single track. The TIMIT is divided into two groups in this experiment, one is comprised of 4620 pieces of voices recorded by 462 persons called TIMIT_1 as training group, the other called TIMIT_2 ?includes 1680 pieces of voices by 168 persons as testing group. Meanwhile,
