New Opportunities with Voice Controlled Accessibility Software: Federated Machine Learning and Linux Mobile Devices

Abstract

For individuals with health limitations, machine learning provides a powerful way to create more useful accessibility software. Voice controlled accessibility software is one such example that allows many to interact with their computers hands free. However, many models related to specialty tasks like accessibility suffer from a lack of data, well trained models, or an ecosystem through which to share them. Machine learning tasks for accessibility software often lack corporate incentives and thus depend upon community-led, manual solutions that potentially compromise privacy. In my paper, I describe a novel way to apply federated learning for voice-based accessibility software. Federated learning allows models to be trained without exposing sensitive voice data to a central server. My software seeks to reduce the complexity of the federated learning process for end users and allows users to use their data generated from existing accessibility software programs. The models produced from the federated learning process can be used to build the backend for new voice control software on mobile devices. Throughout this paper, I describe the technical implementation of this solution and the user experience goals that informed my design. Finally, in the last section of the paper I discuss a new platform, Linux mobile devices, where voice control and general accessibility is lacking. This platform is a useful case study to explore new ways of creating accessibility software. I provide a survey of user interface options on this platform and explain what needs to be done to integrate federated learning models. At the end of the paper, I explain the implications for the future of accessibility software.