Exploring Non-acoustic Factors in ASR and Human Transcription Accuracy of Non-native Speech: The Roles of Word Frequency and Length

Automatic Speech Recognition (ASR) technology is increasingly utilized in second language (L2) learning, offering tools for pronunciation training and intelligibility assessment. While ASR systems have shown to have positive effects on learning, their transcription accuracy for L2 speech may differ from human transcribers and depend on various factors. This study investigated the effects of word frequency and word length on transcription accuracy of French L2 speech, comparing human and ASR transcriptions. Sixteen L2 learners read French texts aloud, which were written down verbatim by 36 first language (L1) French speakers recruited from the general population and three ASR systems. Analyses revealed that transcription accuracy was significantly influenced by word frequency, with more frequent words transcribed more accurately by both humans and ASR. Word length, however, showed no significant effect. These results highlight the role of frequency in transcription accuracy, echoing patterns observed in L1 speech. Future research should examine the interaction between acoustic factors, such as segmental errors, and higher-order non-acoustic factors beyond individual word frequency. Examples of such non-acoustic factors include trigram frequencies or other numerical probabilities that quantify the idiomaticity of a given text.