Phylaudio

Repository for Indo-European language acoustics through space and time, Under review, 2026.

Phylaudio investigates the phylogenetic signal in speech acoustics. Pre-trained audio models (XLS-R, MMS, Whisper, ECAPA-TDNN, openSMILE) extract embeddings from multilingual read speech (FLEURS-R), which are discretised into phylogenetic characters. Maximum-likelihood (IQ-TREE), distance-based (FastME), and Bayesian (BEAST2) methods infer trees over 50 Indo-European languages and compare them against established linguistic phylogenies.

Installation

Prerequisites

All code was developed and tested on Ubuntu 24.04 (linux-64) with CUDA 12+ (used version: 13.2).

Install pixi, a conda-based package manager:

curl -fsSL https://pixi.sh/install.sh | bash

Dependencies

To install the main dependencies, run:

pixi install
pixi run post_install

Additional environments are needed for specific pipeline stages:

pixi install -e regression  # phylogenetic regression (R + brms)
pixi install -e gp          # phylogenetic regression with Gaussian processes (tensorflow, GPflow)
pixi install -e viz         # publication figures (plotly, seaborn)

Usage

See docs/ for detailed instructions for the pipeline scripts. The pipeline runs in six stages:

1. Download — audio, metadata, pre-trained models, reference trees
2. Language identification — embedding extraction, MLP-based LID training, speech quality estimation
3. Sentence trees — per-sentence phylogenetic inference (IQ-TREE or FastME) and supertree construction (ASTRAL)
4. BEAST — Bayesian divergence-time estimation with BEAST2
5. Post-BEAST — NMF population structure, phylogenetic regression, geographic rate surfaces
6. Plots — publication figures

The data/ directory has its own READMEs for subdirectories (eval/, metadata/fleurs-r/, resources/, trees/).

Citation

If you use this code, please cite:

@article{scheidwasser2026phylaudio,
  title   = {Indo-European language acoustics through space and time},
  author  = {Scheidwasser, Neil and Zhu, Harrison Bo Hua and Fosse, Samuel and Huang, Hengguan and Greenhill, Simon J. and Bhatt, Samir and Duch{\^e}ne, David A.},
  year    = {2026},
  note    = {Under review}
}

License

MIT