Developing microphone-based audio commands classifier using convolutional neural network

Abstract

Audio command recognition methods are essential to be recognized for performing user instructions, especially for people with disabilities. Previous studies couldn’t examine and classify the performance optimization of up to twelve audio commands categories. This work develops a microphone-based audio commands classifier using a convolutional neural network (CNN) with performance optimization to categorize twelve classes including background noise and unknown words. The methodology mainly includes preparing the input audio commands for training, extracting features, and visualizing auditory spectrograms. Then a CNN-based classifier is developed and the trained architecture is evaluated. The work considers minimizing latency by optimizing the processing phase by compiling MATLAB code into C code if the processing phase reaches a peak algorithmically. In addition, the method conducts decreasing the frame size and increases the sample rate that is also contributed to minimizing latency and maximizing the performance of processing audio input data. A modest bit of dropout to the input to the final fully connected layer is added to lessen the likelihood that the network will memorize particular elements of the training data. We explored expanding the network depth by including convolutional identical elements, ReLu, and batch normalization layers to improve the network's accuracy. The training progress demonstrated how fast the accuracy of the network is increasing to reach about 98.1 %, which interprets the ability of the network to over-fit the data of training. This work is essential to serve speech and speaker recognition such as smart homes and smart wheelchairs, especially for people with disabilities