| Summary: | Spiking neural network (SNN), largely inspired by a biological neural network, offers
many advantages over the conventional deep neural network (DNN) [1]/ convolutional
neural network (CNN) [2] in energy efficiency and latency due to its event based
processing, which makes it a good candidate for energy efficient voice classification.
Furthermore, the learning mechanism of the SNN typically requires only local
information of pre-synaptic neuron and post-synaptic neuron when a spike happens,
providing a light-weighted energy-efficient and hardware-friendly solution for the
applications of voice recognition and classification.
This dissertation reports a digital spiking neuron based on Leaky Integrate-and-Fire
(LIF) model [3]. As the basic units of neurons in SNN, several LIF models jointly
construct a 3-layer neural network. Subsequently, eight Infinite Impulse Response (IIR)
digital filters ranging from 0 Hz to 400 Hz are designed to analyze human voice in both
time and frequency domains, which extracts the feature of male and female voices. Two
thousands male and female voice clips are used as training sets and five hundred voices
are used as test sets in the neural network. The functionality and performance of the
proposed digital spiking neuron can be verified by test sets to recognize male and
female voice. The obtained results and simulations in MATLAB demonstrate the
superiority of the proposed SNN and determine the potential of such systems in voice
classification.
|
|---|