A Corpus for Large-Scale Phonetic Typology

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

Introduction

Phonetic typology—the study of how speech sounds vary and follow universal patterns across languages—has long been hampered by a massive data bottleneck. Traditionally, researchers had to choose between deep analysis of a handful of languages or shallow analysis of many languages based on previously published (and often inconsistent) reports. The challenge is that creating a phonetically aligned corpus is computationally expensive and usually requires language-specific pronunciation lexicons, which do not exist for the vast majority of the world's languages.

Methodology: The VoxClamantis Pipeline

To bridge this gap, researchers developed VoxClamantis v1.0, a standardized, massive, and multi-lingual dataset of speech segments and their acoustic properties. The researchers built VoxClamantis by processing the CMU Wilderness corpus, which contains audio of New Testament readings.

The VoxClamantis processing pipeline, from raw audio input to phonetic measure extraction.

The pipeline involved three main stages:

  1. Grapheme-to-Phoneme (G2P): Text was converted into phoneme labels using a "universal" G2P system (Unitran) for all languages, supplemented by rule-based tools like Epitran and WikiPron for high-resource languages.
  2. Forced Alignment: Phoneme sequences were mapped to audio. For low-resource languages, reading-specific acoustic models (eHMMs) were trained. For high-resource languages, a pretrained multilingual Automatic Speech Recognition (ASR) model was utilized.
  3. Acoustic Extraction: Standard phonetic measures were automatically extracted using Praat, including formant frequencies (F1–F4) for vowels and spectral properties (peak, center of gravity) for sibilants.

Dataset and Global Coverage

VoxClamantis v1.0 features 635 languages and 690 readings, providing over 400 million aligned segments. This represents a significant leap in scale for phonetic research.

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

The corpus is geo-located and scored by alignment quality using Mel Cepstral Distortion (MCD). While universal G2P methods allow for massive coverage, they introduce some noise, particularly for languages with opaque orthographies, necessitating careful filtering for specific scientific studies.

Key Insights

The scale of VoxClamantis allows for testing long-standing linguistic theories:

  • Phonetic Uniformity: The data supports the idea that if a language has a specific phonological feature, it tends to realize that feature consistently across different sounds (e.g., /i/ and /u/ covarying in height).
  • Dispersion Theory Challenges: Surprisingly, the data did not support a significant relationship between the number of vowels in a language and the acoustic variation (entropy) of those vowels, challenging the idea that languages with more vowels maintain "tighter" acoustic spaces to avoid confusion.

Limitations and Future Work

Despite its scale, the corpus has limitations. The recordings are primarily single-speaker (mostly male) Bible readings, which limits the study of gender-based or inter-speaker variation. Additionally, phoneme labels are estimated from text rather than manually verified, introducing inherent noise. Future work may involve semi-supervised learning to improve lexicons and unsupervised speaker clustering to address the lack of metadata.

Summary updated

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