Acoustic event detection with binarized neural network

Abstract

Implementation of deep learning for Acoustic Event Detection (AED) on embedded systems is challenging due to constraints on memory, computational resources and, power dissipation. Various solutions to overcome this limitation have been proposed. One of the latest methods to overcome this limitation is by using Binarized Neural Network (BNN) which has been proven to achieve approximately 32x memory savings and 58x lower computational resources. XNOR-Net is a type of BNN which uses the XNOR gate to perform a logical function on the input data and give all outputs in binary form. In this project, the XNOR-Net model is constructed and trained for the AED task using urban sound (UrbanSound8K) and bird sound (Xeno-Canto) datasets. Prior to performing the training, the datasets were pre-processed through audio segmentation to produce 1-second sound files. Each audio file is converted from the time domain to Mel-Spectrogram in the frequency domain and thresholding was implemented to convert each spectrogram into a binary image. The images are then reshaped to 32x32 pixels before being used for the training procedure. A performance comparison between BinaryNet and XNOR-Net in terms of the number of hidden layers used was performed and one binary convolutional layer structure XNOR-Net was determined and constructed. The block structure and hyperparameters of the XNOR-Net were analyzed and optimized to achieve a training accuracy of 96.06% and validation accuracy of 94.08%.