Beyond Audio Quality: Understanding and Improving Voice Communication with Low-Resource Deep Learning

Abstract

This paper presents an investigation into the utilization of low-resource deep learning to improve the quality of voice communication in various contexts. The study proposes the creation of Voice Analysis as a Service (VAaaS), which offers spoof detection, interruption detection, voice-to-command generation, and low-resource training to enhance the quality of voice communication. The research addresses four challenges: protecting conversation participants from spoofing attacks, classifying speech overlaps, using English speech for commanding machines, and exploring the use of CPU for training deep learning models. Through this investigation, we aim to provide a comprehensive understanding of how to use low-resource deep learning to facilitate more effective voice interactions between humans and machines.