On Training and Evaluation of Grapheme-to-Phoneme Mappings with Limited Data

Abstract

When scaling to low resource languages for speech synthesis or speech recognition in an industrial setting, a common challenge is the absence of a readily available pronunciation lexicon. Common alternatives are handwritten letter-to-sound rules and data-driven grapheme-to-phoneme (G2P) models, but without a pronunciation lexicon it is hard to even determine their quality. We identify properties of a good quality metric and note drawbacks of naïve estimates of G2P quality in the domain of small test sets. We demonstrate a novel method for reliable evaluation of G2P accuracy with minimal human effort. We also compare behavior of known state-of-the-art approaches for training with limited data. Finally we evaluate a new active learning approach for training G2P models in the low resource setting.

Proceedings of the 18th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology

“…A lot of work has been devoted to the G2P problem (e.g. see ), ranging from those focused on accuracy and model size to those discussing approaches for data-efficient scaling to low resource languages or multilingual modeling Sharma, 2018;.…”

Improved pronunciation prediction accuracy using morphology

“…A lot of work has been devoted to the G2P problem (e.g. see Nicolai et al (2020)), ranging from those focused on accuracy and model size to those discussing approaches for data-efficient scaling to low resource languages or multilingual modeling (Rao et al, 2015;Sharma, 2018;.…”

“…Both non-neural and neural approaches have been studied for transfer learning (Weiss et al, 2016) from a high-resource language for low resource language G2P setting using a variety of strategies including semi-automated bootstrapping, using acoustic data, designing representations suitable for neural learning, active learning, data augmentation and multilingual modeling (Maskey et al, 2004;Davel and Martirosian, 2009;Jyothi and Hasegawa-Johnson, 2017;Sharma, 2018;Ryan and Hulden, 2020;Peters et al, 2017;. Recently, transformer-based architectures have also been used for this task (Engelhart et al, 2021).…”

Developing Pronunciation Models in New Languages Faster by Exploiting Common Grapheme-to-Phoneme Correspondences Across Languages

“…Second, we will describe a multilingual machine-learning G2P model that generalizes to unseen languages. Both approaches significantly reduce the amount of time and effort required to create the G2P component for a new language (see also [7]), which makes it easier to build voice technologies such as ASR systems in more languages.…”