Classifying vocalization from parent-child conversations with a single sensor

Abstract

Recent rapid advances in Machine Learning (ML) have led to many novel applications previously thought infeasible. The proliferation of various speech recognition solutions in various languages shows ML’s promise as the solution for a natural way to interact with devices. Speaker classification has been a subject with less than stellar results, using expensive multi-sensor, multi-device solutions, as well as sacrificing ease of use to achieve acceptable accuracy. We believe utilizing only a single device will greatly enhance the usability, and therefore the utility, of the technology, enabling novel applications such as preventing child abduction and facilitating at-home speech therapy, by automating previously labor-intensive tasks. This study looks at the feasibility of the idea by proposing and evaluating various methods to classify voice as to whether they are from an adult or a small child. This paper demonstrates that specific combinations of existing technologies can now deliver acceptable accuracy in child-adult voice classification without the addition of expensive equipment or cumbersome user interaction, such as training a custom ML model. We believe it is feasible that further research could further improve the technology, facilitating the realization of aforementioned use-cases, even without the aid of a vast and regularized corpus of data.