Project DetailsThis work investigates whether the HuBERT deep-learning model can capture accent-related phonetic differences more precisely than traditional phonetic transcription, while also uncovering additional acoustic and prosodic cues that influence speech intelligibility. In the course of the work, a comparison is made between HuBERT’s internal representations for an L1 speaker (native American English) and an L2 speaker (non-native) and assess how deviations in phonetic transcription correspond to divergences in the model’s latent embeddings.
The findings indicate that the HuBERT model indeed captures information regarding phonetic differences between various accents while considering additional acoustic components that are not included in the standard transcription. These achievements point to significant potential for integrating deep-learning-based models in speech analysis and intelligibility research, and open new avenues for more precise investigation of the impact of accents on listener comprehension.
