Measuring foreign accent strength using an acoustic distance measure

Pronunciations from different speakers are often compared using phonetic transcriptions, even though transcribing speech is time-consuming and error prone. To understand whether this process can be omitted when the goal is to quantify pronunciation differences, we investigate several acoustic-only methods for representing and comparing pronunciations. Specifically, we compute numerical feature representations based on Melfrequency cepstral coefficients and a pre-trained Transformerbased neural network, and make word-level comparisons using Dynamic Time Warping. We use the speech of nonnative and native speakers of English as input to these models, and evaluate the algorithms by comparing their output to human judgements of accent strength. Our results show that the Transformer-based approach outperforms the already wellperforming transcription-based method, while being minimally affected by individual speaker differences. These results suggest that phonetically transcribing speech is not necessary to quantify pronunciation differences when a pre-trained Transformerbased neural model is available.