duncan@convex.csd.uwm.edu (Shan D Duncan) (01/11/91)
[Try this again] I came across a reference of some work done at Sea World/Hubbs Research (1988?) about using NN to classify Killer Whale vocalizations. I remember reading an issue of AI that also mentioned this work but no real details were given. Any information available? I assume that digitized sonagrams/spectragrams were used and then pattern recognition of the time/frequency picture. I would like to know if the digitized data from an A/D board could be used directly (waveform) without a transformation or if a matrix of amplitude values at specific frequencies over time is necessary via FFT and windowing. I would like to train a network on species specific vocalizations and then use it to classify. I would ultimately like to use the same principle to handle vocalizations on an individual level where the amount of data makes the traditional approach difficult (i.e. recording the vocalization, obtaining a frequency/time representation, measuring variables then using statistical technique both univariate and multivariate to obtain a measure of variability and similarity/dissimilarity). I am not sure if I am really asking the proper questions or if NNs are really the appropriate technique but it seems made for a situation where one must handle voluminous data and the resulting signal could be rather complex with both Frequency Modulation and Amplitude Modulation. Please assume a very basic level of understanding, I do not mind being told things I already know. :-) Thank you for any help or information, programs (unix), -Shan Duncan Dept. of Biological Sciences University of Wisconsin--Milw. Milwaukee, WI. 53201