Desarrollo de una aplicación móvil que permita clasificar sonidos para personas con discapacidad auditiva utilizando deep learning.

Abstract

Over the years, cell phones have been evolving to provide their users with different functionalities no matter what their field or need is, becoming an indispensable day-to-day tool, due to their ability to offer great utility. People with reduced capacities are able to communicate many times thanks to these devices, helping them with simple tasks such as communicating with others, even asking for emergency services. All thanks to the dynamism offered by these devices. Over time there has been a growing interest in artificial intelligence, its uses are limited to the imagination of the people who put them to use, allowing people to automate tedious and time-consuming processes that would take too long if done manually. Since cell phones have evolved to the point of being considered "pocket phones", the idea is to create solutions that use artificial intelligence to solve people's problems, without having to access large servers or web pages constantly, everything can be achieved with applications that can be installed on a phone. In order to support people through technology, it is planned to build an application for smartphones, capable of recognizing sounds around and present a result of what is heard, using artificial intelligence tools, focusing on people with hearing difficulties, in order to recognize the sounds of the environment that can not perceive correctly. This project was carried out with tools such as Android Studio, Teachable Machine, which uses as a base package the Command Speech Recognizer algorithm of Tensorflow, which has a workflow capable of identifying isolated words and forming a knowledge network based on these, also working with sounds, so it is perfectly suited to the project being carried out, which is why it will be exported in Tensorflow Lite. During the construction of the model, the aim is to test the neural network in order to evaluate its accuracy, placing an acceptable index to meet a level of satisfaction, which will be measured using the confusion matrices for each of the domains that have been defined in the model and determine whether the model is suitable to work in a real-world case. The way in which the application works using the trained model will be that by means of the device's microphone it will be able to listen to what is in the environment and will be able to deliver a matching result depending on the sounds that have been recognized in the environment and are in a domain of the model. Thanks to the tests that were performed in a real environment, it can be concluded that the model has a high level of acceptance, with an accuracy just to the necessary average that was proposed to be considered as suitable for the trained model, so the model is considered suitable for application in a real world case.