VoxCommunis: A Corpus for Cross-linguistic Phonetic Analysis

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5286–5294 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5286 VoxCommunis: A Corpus for Cross-linguistic Phonetic Analysis Emily P. Ahn, Eleanor Chodroff University of Washington, University of York Washington, USA, York, UK eahn@uw.edu, eleanor.chodroff@york.ac.uk Abstract Cross-linguistic phonetic analysis has long been limited by data scarcity and insufficient computational resources. In the past few years, the availability of large-scale cross-linguistic spoken corpora has increased dramatically, but the data still require considerable computational power and processing for downstream phonetic analysis. To facilitate large-scale cross-linguistic phonetic research in the field, we release the VoxCommunis Corpus, which contains acoustic models, pronunciation lexicons, and word- and phone-level alignments, derived from the publicly available Mozilla Common Voice Corpus (Ardila et al., 2020). The current release includes data from 36 languages. The corpus also contains acoustic-phonetic measurements, which currently consist of formant frequencies (F1–F4) from all vowel quartiles. Major advantages of this corpus for phonetic analysis include the number of available languages, the large amount of speech per language, as well as the fact that most language datasets have dozens to hundreds of contributing speakers. We demonstrate the utility of this corpus for downstream phonetic research in a descriptive analysis of language-specific vowel systems, as well as an analysis of “uniformity” in vowel realization across languages. The VoxCommunis Corpus is free to download and use under a CC0 license at https://osf.io/t957v/wiki/home/. Keywords: phonetics, typology, G2P, alignment, speech 1. Introduction For a thorough understanding of cross-linguistic pho- netic variation and systematicity, big data from a di- verse set of languages is necessary. In 2009, it was noted that despite the advances made in speech tech- nology and computational power, there had been “sur- prisingly little change in style and scale of [phonetic] research” from 1966 onwards (Liberman, 2009). Since even that period, considerable advances have been made for increased processing power of large-scale phonetic corpora, but largely within a single “high- resource” language such as English, in which relevant data for automatic speech processing approaches al- ready exist. Until recently, the movement towards large-scale, cross-linguistic phonetic research has been somewhat limited. Previous large-scale cross-linguistic phonetic studies have mostly been meta-analyses that rely on a standardized phonetic measure and a plethora of published cross-linguistic research (Whalen and Levitt (1995) for vowel f0, Becker-Kristal (2010) for vowel F1 and F2, Chodroff et al. (2019) for stop VOT). Prior to 2020, existing multilingual speech corpora contained maximally twenty-some languages (Harper, 2011; Schultz et al., 2013), or insufficient data for most phonetic analyses (Ladefoged and Maddieson, 2007). More recently, Salesky et al. (2020) presented the VoxClamantis Corpus for large-scale phonetic ty- pology that provided phonetic data for over 500 lan- guages, based on recordings and transcripts from the CMU Wilderness Corpus (Black, 2019). Only in the past few years (e.g., from 2019) have these massively multilingual corpora been made publicly available (see Section 2 for an overview). In addition, the tools nec- essary to …

Content truncated to 500 words