Grapheme-to-phoneme conversion (g2p) is the task of predicting the pronunciation of words from their orthographic representation. Historically, g2p systems were transition- or rule-based, making generalization beyond a monolingual (high resource) domain impractical. Recently, neural architectures have enabled multilingual systems to generalize widely; however, all systems to date have been trained only on spelling-pronunciation pairs. We hypothesize that the sequences of IPA characters used to represent pronunciation do not capture its full nuance, especially when cleaned to facilitate machine learning. We leverage audio data as an auxiliary modality in a multi-task training process to learn a more optimal intermediate representation of source graphemes; this is the first multimodal model proposed for multilingual g2p. Our approach is highly effective: on our in-domain test set, our multimodal model reduces phoneme error rate to 2.46%, a more than 65% decrease compared to our implementation of a unimodal spelling-pronunciation model-which itself achieves state-of-the-art results on the Wiktionary test set. The advantages of the mullimodal model generalize to wholly unseen languages, reducing phoneme error rate on our out-of-domain test set to 6.39% from the unimodal 8.21%. a more than 20% relative decrease. Furthermore, our training and test sets are composed primarily of low-resource languages, demonstrating that our multimodal approach remains useful when training data are constrained.
展开▼
