ICS2001-05-02: An artificial neural network for recognition of simulated dolphin whistles - T. Burgess

It is known that dolphins are capable of understanding 200 "word" vocabularies with sentence complexity of three or more "words", where words consist of audio tones or hand gestures. An automated recognition method of words where a word is a defined whistle, within a predetermined acceptable degree of variance, could allow words to be both easily reproducible by dolphins and identifiable by humans. We investigate a neural network to attempt to distinguish four artificially generated whistles from themselves and from common underwater environmental noises, where a whistle consists of four variations of a fundamental whistle style. We play these whistle variations into the dolphins normal tank environment and then record from a separate tank hydrophone. This results in slight differences for each whistle variation's spectrogram, the complete collection of which we use to form the neural network training set. For a single whistle variation, the neural network demonstrates strong output node values,greater than 0.9 on a scale of 0 to 1. However, for combinations of"words", the network exhibits poor training performance and an inability to distinguish between words. To validate this, we used a test set of 41 examples, of which only 22 were correctly classified. This result suggests that an appropriately trained back propagation neural network using spectrographic analysis asinputs is a viable means for a very specific whistle recognition,however a large degree of whistle variation will dramatically lower theperformance of the network, past that required for acceptable recognition.

ICS2001-05-02.pdf — PDF document, 593Kb