Phone Inventories and Recognition for Every Language

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1061–1067 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1061 Phone Inventories and Recognition for Every Language Xinjian Li, Florian Metze, David R. Mortensen, Alan W. Black, Shinji Watanabe Carnegie Mellon University Language Technologies Institute 5000 Forbes Ave, Pittsburgh, PA 15213 US {xinjianl, fmetze, dmortens, awb, swatanab}@andrew.cmu.edu Abstract Identifying phone inventories is a crucial component in language documentation and the preservation of endangered languages. However, even the largest collection of phone inventory only covers about 2000 languages, which is only 1/4 of the total number of languages in the world. A majority of the remaining languages are endangered. In this work, we attempt to solve this problem by estimating the phone inventory for any language listed in Glottolog, which contains phylogenetic information regarding 8000 languages. In particular, we propose one probabilistic model and one non-probabilistic model, both using phylogenetic trees (“language family trees”) to measure the distance between languages. We show that our best model out- performs baseline models by 6.5 F1. Furthermore, we demonstrate that, with the proposed inventories, the phone recognition model can be customized for every language in the set, which improved the PER (phone error rate) in phone recognition by 25%. Keywords: Phone Inventory, Multilingual Speech Recognition, Endangered Languages, Language Documentation 1. Introduction A fundamental aspect of the description or documen- tation of any language is establishing its phone inven- tory (Bird and Simons, 2003; Michaud et al., 2018). This is a necessary prerequisite to further phonetic and phonological analysis (including transcribing text, dis- covering allophonic patterns, and developing an or- thography), these are foundations upon which other facets of linguistic description can be built. Tradition- ally, phone inventories have been discovered by field linguists using a mixture of audio, visual, and lexical tools to arrive at a set of sounds sufficient to char- acterize the phonetics of the language. The largest collection of phone inventory aggregated so far is the PHOIBLE dataset (Moran, Steven and McCloy, Daniel and Wright, Richard, 2014), which is a collection of phone inventories from over 2000 languages. How- ever, there are around 8000 languages in the world, for most of which no documented phone inventory ex- ists. Unfortunately, those languages are typically en- dangered (Nettle et al., 2000). Language preserva- tion projects typically target languages in this category. Field linguists starting work on a new language will benefit from knowing, in approximate terms, what the phone inventory of that language is like. In this work, we attempt to solve this problem by esti- mating the phone inventory for any language listed in Glottolog (Nordhoff and Hammarström, 2011), which contains around 8000 languages. In particular, we take advantage of the phylogenetic trees from Glottolog (as this information is available for almost every language in the world)1. We propose two approaches that ex- 1Since linguists do not always agree upon the phyloge- netic groupings of languages—especially of poorly-studied languages—the trees from Glottolog are necessarily …

Content truncated to 500 words