The rise of audio-as-data in social science research accentuates a fundamental challenge: establishing reproducible and reliable methodologies to guide this emerging area of study. In this study, we focus on the reproducibility of audio-as-data preparation methods in computational communication research and evaluate the accuracy of popular audio-as-data tools. We analyze automated transcription and computational phonology tools applied to 200 episodes of conservative talk shows hosted by Rush Limbaugh and Alex Jones. Our findings reveal that the tools we tested are highly accurate. However, despite different transcription and audio signal processing tools yield similar results, subtle yet significant variations could impact the findings’ reproducibility. Specifically, we find that discrepancies in automated transcriptions and auditory features such as pitch and intensity underscore the need for meticulous reproduction of data preparation procedures. These insights into the variability introduced by different tools stress the importance of detailed methodological reporting and consistent processing techniques to ensure the replicability of research outcomes. Our study contributes to the broader discourse on replicability and reproducibility by highlighting the nuances of audio data preparation and advocating for more transparent and standardized practices in this area.
Read full abstract