Speaker recognition in tactical communications
Richard Ricart, J. Cupples, Laurie H. Fenstermacher · 2002
Tactical communications are inherently short and exhibit a great deal of channel variability. A novel speaker recognition technique is described in which on-line training is utilized to circumvent the need for excessive speaker or channel modeling. The technique incorporates both feature set fusion and classifier fusion. Separate classifiers are trained for each feature set: LPC cepstra with and without RASTA filtering concomitant with delta and acceleration cepstra. The results of the individual are then adjudicated to the correct speaker. The speaker recognition algorithm was baselined with the KING database, used extensively in speaker recognition. A subsequent evaluation, conducted on the Rome Laboratory GREENFLAG tactical communications database, resulted in 93% correct identification of 41 speakers.>