Disign of a workbench for audio analysis

Abstract

Acoustic Scene Classification (ASC) is the task to identify audio recordings into one of several predefined acoustic scene classes. There is a huge amount of systems that have been developed to tackle the ASC problem using different machine learning-based techniques. Therefore, building a program that combines multiple techniques has become a necessary need to properly assess the effects of these techniques on the performance of the ASC systems and compare between it in an easy way. Motivated by these needs, we have designed in our work an acoustic scene classification-based workbench that consists of variety of techniques and methods related to this task. The goal of this work is to allow the users to apply the techniques they are desiring to use on a chosen dataset, make changes in the parameters fast, create big tests, and visualize the results of these tests. To accomplish this task, we have created a workbench that is composed of modules consisting various tools for each module. Specifically, we have implemented 5 modules concerning the ASC task stages which are respectively: Data loading and visualization, for importing a Dataset and get an overview of it. Data transformation, for extracting relevant features from the dataset. Data augmentation, for increasing data size. Training, for train the employed Deep Neural Networks on the obtained dataset and evaluate its performance. And lastly Prediction, for evaluating the performance of the trained models on making an accurate class prediction of a chosen audio recording. Thereafter, for executing our workbench to see how it performs, we have used a modified TAU Urban Acoustic Scenes 2019 dataset that consist of 7192 audio recordings with 10 various classes on 5 different cities for each, and tested it on different existing techniques while recording the obtained accuracy on a table of results. Keywords: Acoustic scene classification, Workbench, Machine learning, Data Augmentation, Deep Neural Network.