Online testing for assessing speech intelligibility

Abstract

The use of online testing in the behavioral sciences is increasing due to the potential for lower cost and faster completion than traditional in-person laboratory testing. Online testing presents a special challenge for speech research due to the variety of listeners’ acoustic environments. For example, listeners may use headphones or speakers of various types, complete the task in different levels of background noise, or vary in their hearing abilities. Here we presented spoken sentences in speech-shaped noise to participants from the United States recruited online using Amazon Mechanical Turk at SNRs of -2, -5, and -8 dB (n≥50 per SNR). We compared these online results to normative data collected using these same sentences in the lab (SNR of -5 dB; n = 30). Preliminary results suggest a reasonable correspondence between average intelligibility scores for individual sentences presented using the two methods (Spearman rho = 0.79). Standard deviations suggest greater variability in the online responses (average SD across sentences = 1.24) relative to the lab responses (average SD = 0.86). We conclude that despite increased variability, online testing can provide reasonable ratings of intelligibility and in many circumstances is a feasible method to supplement or replace laboratory testing.