Homograph Disambiguation with Text-to-Text Transfer Transformer

In recent years, the Text-to-Text Transfer Transformer (T5) neural model has proved very powerful in many text-to-text tasks, including text normalization and grapheme-to-phoneme conversion. In the presented paper, we fine-tuned the T5 model for the task of homograph disambiguation, which is one of the essential components of text-to-speech (TTS) systems. To compare our results to those of other studies, we used an online dataset of US English homographs called Wikipedia Homograph Data. We present our results, which outperformed the previously published single-model approaches. We also focus on more detailed error analysis, model performance on different types of homographs, and the impact of training set size on homograph disambiguation.