Biomimetic spike-based algorithms and hardware for sound classification, localization, and speech recognition

The objective of the thesis work is to design real-time, spike-based algorithms and implementation for biomimetic sound processing systems. An acoustic direction finding (ADF) system was designed and implemented in hardware to process transient sounds. Additionally, a top-down attentional mechanism model based on a study of mammalian brain activity was designed and explored to improve speech recognition. The front-end of the ADF system, which mimics a mammalian peripheral auditory system, generates spiking neuron firings as its output. The back-end algorithm was developed in MATLAB and an FPGA-based neural network was its final embodiment. The algorithm accomplishes sound detection, classification, direction finding, and localization of various kinds of audio data under noisy conditions and from reverberant environments. The attentional model was integrated with the front-end processing to help segregate a target sound source from masker sound sources as well as improving the classification accuracy of the target source.;The neural-network-based sound classification and localization algorithm was first developed and tested using weaponry sound data obtained in the field. The algorithm is able to differentiate and trace various gunfire acoustic signatures in the presence of high background noise. The algorithm can locate the sound source by using single or multiple microphone array sites. Compared to a least square time difference of arrival algorithm, the neural-network-based algorithm has higher detection rate and more accurate localization. The complete back-end processing system was implemented on a single Xilinx Virtex-5 FPGA chip. The neural-network-based algorithm was also modified for a frog habitat monitoring application to demonstrate that the algorithm can be useful for applications other than weaponry classification and localization.;Literature was reviewed and a functional, biologically-based, top-down attentional model was formulated, coded, and tested using speech signals with varying target masker ratios. The model improves the correctness of word identification of target speech by up to 50% in a noisy environment when the masker source is either a white noise signal or a speech- like signal.;The thesis work presents the first spike-based transient sound classification and localization algorithm using neural networks, the first spike-based frog habitat monitoring algorithm, and a novel top-down, biologically-based attentional model.