A Corpus for Large-Scale Phonetic Typology

Elizabeth Salesky, Eleanor Chodroff, Tiago Pimentel, Matthew Wiesner, Ryan Cotterell, Alan W. Black, Jason Eisner

Abstract

A major hurdle in data-driven research on typology is having sufficient data in many languages to draw meaningful conclusions. We present VoxClamantis v1.0, the first large-scale corpus for phonetic typology, with aligned segments and estimated phoneme-level labels in 690 readings spanning 635 languages, along with acoustic-phonetic measures of vowels and sibilants. Access to such data can greatly facilitate investigation of phonetic typology at a large scale and across many languages. However, it is non-trivial and computationally intensive to obtain such alignments for hundreds of languages, many of which have few to no resources presently available. We describe the methodology to create our corpus, discuss caveats with current methods and their impact on the utility of this data, and illustrate possible research directions through a series of case studies on the 48 highest-quality readings. Our corpus and scripts are publicly available for non-commercial use at https://voxclamantisproject.github.io.

Introduction

Phonetic typology—the study of how speech sounds vary and what patterns are universal across the world's languages—has long been a "data-poor" field. Historically, researchers faced a trade-off between depth (high-quality data for few languages) and breadth (inconsistent measures from disparate studies). While the CMU Wilderness corpus provided raw audio for over 600 languages, it lacked the precise phoneme-level alignments necessary for phonetic research. VoxClamantis v1.0 bridges this gap by processing raw data into a standardized, research-ready format.

Methodology

The researchers developed a multi-stage pipeline to transform raw Bible readings into a structured phonetic corpus. The process begins with segmenting audio into utterances using text-to-speech alignments. Next, a Grapheme-to-Phoneme (G2P) stage converts text to phonetic representations using tools like Unitran, Epitran, or WikiPron. Forced alignment is then performed using an eHMM (embedded Hidden Markov Model) to determine exact phoneme boundaries. Finally, acoustic features such as vowel formants (F1–F4) and sibilant spectral moments are extracted using Praat.

The multi-stage pipeline for transforming raw audio into aligned phonetic segments and acoustic measures.

The VoxClamantis v1.0 Dataset

The resulting corpus includes 690 readings across 635 languages, providing standoff markup for over 400 million labeled segments. To ensure data quality, the authors used Mel Cepstral Distortion (MCD) to score alignment accuracy across the global dataset, identifying a high-quality subset for initial case studies.

Global distribution of the 635 languages in VoxClamantis v1.0, colored by Mel Cepstral Distortion (MCD) alignment quality.

Key Insights and Results

Case studies on the highest-quality readings yielded several significant findings:

  • Phonetic Uniformity: Data suggests languages "yoke" similar sounds together. For instance, a high-pitched realization of /i/ often correlates with a high realization of /u/, suggesting universal structural forces in sound system organization.
  • Dispersion Myth: Contrary to the hypothesis that larger vowel inventories lead to greater acoustic spacing to avoid confusion, the authors found no significant correlation between inventory size and acoustic entropy.
  • Scalability: The study demonstrates that large-scale automated alignments, even if "noisy," can replicate findings from smaller manual studies, proving that scale can compensate for algorithmic error.

Limitations and Future Work

The corpus is currently limited by speaker diversity, as Bible readings are typically single-speaker and non-dramatic. Additionally, the universal G2P approach (Unitran) can introduce systematic errors in languages with opaque orthographies. Future work will focus on unsupervised speaker clustering for multi-speaker recordings and semi-supervised learning to improve G2P accuracy for low-resource languages.

The corpus and scripts are available at voxclamantisproject.github.io.

Summary updated

Abstract

A major hurdle in data-driven research on typology is having sufficient data in many languages to draw meaningful conclusions. We present VoxClamantis v1.0, the first large-scale corpus for phonetic typology, with aligned segments and estimated phoneme-level labels in 690 readings spanning 635 languages, along with acoustic-phonetic measures of vowels and sibilants. Access to such data can greatly facilitate investigation of phonetic typology at a large scale and across many languages. However, it is non-trivial and computationally intensive to obtain such alignments for hundreds of languages, many of which have few to no resources presently available. We describe the methodology to create our corpus, discuss caveats with current methods and their impact on the utility of this data, and illustrate possible research directions through a series of case studies on the 48 highest-quality readings. Our corpus and scripts are publicly available for non-commercial use at https://voxclamantisproject.github.io.
BibTeX
@inproceedings{salesky2020corpus,
  author = {Elizabeth Salesky and Eleanor Chodroff and Tiago Pimentel and Matthew Wiesner and Ryan Cotterell and Alan W. Black and Jason Eisner},
  title = {{A Corpus for Large-Scale Phonetic Typology}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
  doi = {10.18653/v1/2020.acl-main.415},
}
Figure