Multi-Label Classification of Voice Calls from Power Plant Operation Centers

Abstract

This paper presents a machine learning-based pipeline to classify, in six labels, voice calls from power plant operation centers from ENGIE Brasil Energia (private power producer). The pipeline consists of a customized speech-to-text model from Amazon Web Services (AWS) followed by a multi-label text classification model. Our experiments showed how to leverage the performance of Amazon Transcribe with a custom vocabulary, as well as the predictive performance for different tree-based machine learning models. The work aims to facilitate the audit of internal actions and increase operational efficiency from the post-operation activities from ENGIE Brasil Energia (EBE). We achieved great predictive results, high accuracy, and Fl-score for all six labels.