Over the last 20 years, automatic speech recognition (ASR) has improved substantially and spread into people’s everyday lives through hands‐free texting and intelligent personal assistants. ASR has also gained recognition within the field of second language learning as the transcript provided by ASR programs can support noticing of pronunciation errors. Given the growing interest and research in ASR, it is important to examine the accuracy of the technology for nonnative speech, especially given that studies have documented that second language speakers have been frustrated by low levels of ASR recognition. Using a three‐way ANOVA, this study examines the transcription accuracy of two ASR‐dictation programs, Google and Windows Speech Recognition ( WSR ), to understand accuracy as compared to human listeners for learners of three different first language backgrounds (Arabic, Chinese, and Spanish) on two different types of tasks (free and read speech). The results show that while Google had a higher level of transcription accuracy than WSR , WSR actually showed greater correlation with human listener intelligibility. While the speakers’ language background was not shown to have a significant effect, the task did, with WSR losing substantial accuracy in the free speech task. The results show a need for continued development and advancement of ASR technologies for second language speech.
McCrocklin et al. (Thu,) studied this question.