Spike Encoding of Mel Frequency Cepstrum as an Input Stage for Speech Recognition

Spiking Neural Network (SNN) uses spike which is a discrete event that occurs in time, as opposed to continuous time signals. This allows low cost resources implementation in electronic devices because the spikes occurs in particular time events. In this research, we implement and test various spike encoding techniques of mel frequency cepstrum designed for automatic speech recognition system with SNN method. The implementation is done in the Xilinx Pynq-Z2 Field Programmable Gate Array (FPGA). This Pynq-Z2 is equipped with an ARM Processor that will be used as the whole computational process. The preprocessing audio signal uses Mel Spectogram method which is the output spectrogram and translated into spike trains the will serve as input to a SNN. Results show that the latency spike encoding method results in the least amount of spikes. Substantial memory and power savings from spike encoding of audio signals has been shown and can potentially be used as the input stage for a low-cost hardware implementation of a speech recognition system.