Examining the Impact of Dialectal Variation on Speech-to-Text Algorithm Fairness

Abstract

This research aims to investigate how dialectal variation—a shift in language caused by physical and social environmental influences—impacts the fairness of speech-to-text algorithms. A surge in dependence on these algorithms has resulted from the proliferation of speech-to-text technology in recent years in copious applications including automated captioning, virtual assistants, speech recognition software, and court transcription. However, there is a growing concern that these algorithms may not perform equally well for speech from different dialects, which can result in bias and inaccuracies. This study focuses on comparing the performance of speech-to-text algorithms when transcribing speech from different dialects, specifically examining the effect of dialectal variation on accuracy and error rate. The study uses a sample of speech from multiple dialects, including but not limited to British English, Southern American English, and Pacific Northwest English. The study investigates the extent to which dialectal variation affects the fairness of speech-to-text algorithms for different demographic groups, notably gender. The methodology of this research involves the use of word error rate (WER) to evaluate the performance of the speech-to-text algorithms. The study also uses qualitative methods, such as manual transcription and annotation, to provide a more in-depth understanding of specific errors and biases that may be present in the algorithm’s output. The findings of this research have important implications for the development and use of speech-to-text technology. By identifying and addressing sources of bias in speech-to-text algorithms that may be caused by dialectal variation, this study aims to contribute to the development of more fair and accurate speech-to-text algorithms for a wide range of dialects. Moreover, this study can help with the design of speech-to-text technology to better serve and accommodate diverse communities, including those who speak non-standard dialects. In summary, this research contributes to a greater knowledge of how dialectal variety influences the fairness of speech-to-text algorithms, which is essential for the development of more equitable and inclusive technologies. The findings from this study can sanction the design of speech-to-text algorithms to better serve diverse communities and raise the overall functionality of speech-to-text technology.