AlloVera: A Multilingual Allophone Database

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5329–5336 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5329 AlloVera: A Multilingual Allophone Database David R. Mortensen∗, Xinjian Li∗, Patrick Littell†, Alexis Michaud‡, Shruti Rijhwani∗, Antonios Anastasopoulos∗, Alan W. Black∗, Florian Metze∗, Graham Neubig∗ ∗Carnegie Mellon University; †National Research Council of Canada; ‡CNRS-LACITO ∗5000 Forbes Ave, Pittsburgh PA 15213, USA; †1200 Montreal Rd, Ottawa ON K1A0R6, Canada; ‡7 rue Guy Môquet, 94800 Villejuif, France ∗{dmortens, xinjianl, srijhwan, aanastas, awb, fmetze, neubig}@cs.cmu.edu; †patrick.littell@nrc-cnrc.gc.ca; ‡alexis.michaud@cnrs.fr Abstract We introduce a new resource, AlloVera, which provides mappings from 218 allophones to phonemes for 14 languages. Phonemes are contrastive phonological units, and allophones are their various concrete realizations, which are predictable from phonological context. While phonemic representations are language specific, phonetic representations (stated in terms of (allo)phones) are much closer to a uni- versal (language-independent) transcription. AlloVera allows the training of speech recognition models that output phonetic transcriptions in the International Phonetic Alphabet (IPA), regardless of the input language. We show that a “universal” allophone model, Allosaurus, built with AlloVera, outperforms “universal” phonemic models and language-specific models on a speech-transcription task. We explore the implications of this technology (and related technologies) for the documentation of endangered and minority languages. We further explore other applications for which AlloVera will be suitable as it grows, including phonological typology. Keywords: Phonology, Allophones, Database, Automatic Speech Recognition 1. Introduction Speech can be represented at various levels of abstraction (Clark et al., 2007; Ladefoged and Johnson, 2014). It can be recorded as an acoustic signal or an articulatory score. It can be transcribed with a panoply of detail (a narrow tran- scription), or with less detail (broad transcription). In fact, it can be transcribed retaining only those features that are contrastive within the language under description, or with abstract symbols that stand for contrastive units. This latter mode of representation is what is called a phonemic rep- resentation while the finer-grained range of representations are phonetic representations. Most NLP technologies that represent speech through transcription do so at a phonemic level (that is, words are represented as strings of phonemes). For language-specific models and questions, such repre- sentations are often adequate and may even be preferable to the alternatives. However, in multilingual models, the language-specific nature of phonemic abstractions can be a liability. The added phonetic realism of even a broad pho- netic representation moves transcriptions closer to a univer- sal space where categories transcend the bounds of a partic- ular language. This paper describes AlloVera1, a resource that maps be- tween the phonemic representations produced by many NLP tools—including grapheme-to-phoneme (G2P) transducers like our own (Mortensen et al., 2018)—and broad phonetic representations. Specifically, it is a database of phoneme- allophone pairs (where an allophone is a phonetic realiza- tion of a phoneme—see § 1.1. below) for 14 languages. It is designed for notational compatibility with existing G2P systems. The phonetic representations are relatively broad, a consequence of our sources, but …

Content truncated to 500 words