Speech command classification based on deep neural networks

Abstract

In our day-to-day life there are lots of sounds that we are processing. To process these sounds our brain absorb sound signals and provide us informative knowledge. For human being this is not possible to extract every sounds properly so that, there are lots of equipment which helps us to extract essential information from an audio source. Around the year lots of model came to help thorough extract informations using various algorithms. Also, some models are Convolutional Neural Network (CNN), Region-Convolutional Neural Network (R-CNN), Artificial Neural Network (ANN), VGG16, ResNet50 and Numerous machine learning algorithms have been utilized to effectively categorize audio, and these methods have recently demonstrated encouraging results in separating spectrotemporal images from various sound classifications. The study purpose of this research was to analyze which feature extraction method shows maximum result using Convolutional Neural Network (CNN), VGG16 and ResNet50. In the proposed model, MFCC feature extraction method are taken from the dataset and trained using a multiple layer-based con volution neural network. In the experimental assessment, a sound dataset consisting of 105829 audio clips separated up into multiple groups of important sounds during study used to develop the models. Additionally, we evaluated the models’ validity which reach an accuracy of 94.53% on Speech Command dataset.