Statistical considerations for crowdsourced perceptual ratings of human speech productions

Crowdsourcing has become a major tool for scholarly research since its introduction to the academic sphere in 2008. However, unlike in traditional laboratory settings, it is nearly impossible to control the conditions under which workers on crowdsourcing platforms complete tasks. In the study of communication disorders, crowdsourcing has provided a novel solution to the collection of perceptual ratings of human speech production. Such ratings allow researchers to gauge whether a treatment yields meaningful change in how human listeners' perceive disordered speech. This paper will explore some statistical considerations of crowdsourced data with specific focus on collecting perceptual ratings of human speech productions. Random effects models are applied to crowdsourced perceptual ratings collected in both a continuous and binary fashion. A simulation study is conducted to test the reliability of the proposed models under differing numbers of workers and tasks. Finally, this methodology is applied to a data set from the study of communication disorders.