Detection of mobile phone interference in environmental audio recordings

Abstract

The era of machine learning has been beginning to be an engine for the development and creation of applications for a few years and the public is not aware that machine learning is on most of the technological devices on the market. Nowadays, this technology is attracting a lot of attention to the researches and it is giving results that were not possible before: the development of vehicles without drivers to recognize a traffic signal, text detection for translation or the recognition of voice and sounds etc. This kind of techniques are made possible by machine learning. Machine learning consists in teaching computers to do what is natural for people: learn through examples. Therefore, it is necessary to have available a large amount of data to provide these examples. The purpose of this thesis is to develop an application that detects the interference produced by mobile phones in audio recordings through a deep learning architecture, known as feedforward neural networks (FNN), which is used in many machine learning methods. These neural networks will carry out the necessary learning to analyze an acoustic signal and differentiate whether a test audio example contains interference. To perform this learning, first the sound file is represented in the frequency domain through the mel spectrogram. Deep neural networks (DNN), use a layered structure of units to extract characteristics of the given sound representation input with an increased abstraction in each layer. This increases the ability of the network to efficiently learn the highly complex relationship between sound representation and target sounds. In this thesis, we will classify interference and no interference categories by constructing a simple model of these samples as inputs and the binary classification at the output.