Phylaudio

Installation

Dependencies

pixi install
pixi run post_install

[!WARNING] Embedding extraction and LID training were run on a CUDA 12.6 GPU; higher CUDA versions are likely to work but are untested. A GPU is not required to install the environment or to run the downstream phylogenetic analyses — on CPU-only machines, pixi install will resolve the CPU build of PyTorch automatically.

Data

Download pipeline inputs:

pixi run download_models           # download pre-trained audio models
pixi run download_fleurs           # download FLEURS-R audio dataset
pixi run download_glottolog        # extract lineages from FLEURS-R
pixi run download_reference_trees  # extract and process reference trees
pixi run download_geojson          # download language polygon data (Glottography)

Download external data from Zenodo (see ZENODO.md for full manifest):

# From the repo root:
tar -xzf phylaudio_zenodo.tar.gz

This unpacks BEAST2 posteriors, XLS-R embeddings, and regression outputs into data/.

Language identification

By default, training metrics are written to a local CSV file via Lightning's CSVLogger. To use Weights & Biases instead, set up an account (see the Quickstart) and pass --project <name>.

# CSV logger (default, no account required)
pixi run lid --dataset fleurs-r --model_id NeMo_ambernet

# wandb logger
pixi run lid --dataset fleurs-r --model_id NeMo_ambernet --project phylaudio

Sentence-wise trees

Distance trees

pixi run sentence_distance --dataset fleurs-r --model_id NeMo_ambernet --ebs 1

Discrete trees

pixi run sentence_discrete --dataset fleurs-r --model_id NeMo_ambernet

ASTRAL species tree estimation

pixi run sentence_astral pdist
pixi run sentence_summary pdist

BEAST2

Standard run

pixi run beast2 -beagle_SSE -threads 8 -seed 889 data/trees/beast/speech/0.01_brsupport/input.xml

Prior only

pixi run beast2 -sampleFromPrior -beagle_SSE -threads 8 -seed 889 data/trees/beast/speech/0.01_brsupport/prior.xml

Combining runs

scripts/beast_combine_logs.sh data/trees/beast/speech/0.01_brsupport input_v12

Tree summaries (CCD)

pixi run treeannotator -topology CCD0 data/trees/beast/speech/0.01_brsupport/input_combined_resampled.trees input_combined_resampled.ccd0

Phylogenetic networks

pixi run network_analysis data/trees/beast/speech/0.01_brsupport/input.xml

Phylogenetic regression

Install the regression environment:

pixi install -e regression

Required files

Before running regression or plotting, the following files must be present:

File	Source
data/trees/beast/speech/0.01_brsupport/input_combined_resampled.mcc	Zenodo (speech MCC tree)
data/trees/beast/speech/0.01_brsupport/input_combined_resampled.log	Zenodo (speech BEAST log)
data/trees/beast/speech/0.01_brsupport/input_combined_resampled.trees	Zenodo (speech posterior trees)
data/trees/beast/speech/0.01_brsupport/prior_1.log	Zenodo (speech prior log)
data/trees/references/raw/iecor.nex	pixi run download_reference_trees (IECoR MCC tree)
data/trees/beast/iecor/raw.trees	pixi run download_reference_trees (IECoR posterior)
data/trees/beast/iecor/raw.log	pixi run download_reference_trees (IECoR posterior log)
data/trees/beast/iecor/prior/raw.log	pixi run download_reference_trees (IECoR prior log)
data/trees/beast/iecor/prunedtomodern.trees	pixi run download_reference_trees (auto-pruned)

Prepare regression data

Generates metadata CSVs (with and without phoneme inventory) for both speech and cognate trees. Reads MCC trees from data/trees/beast/:

pixi run -e regression prepare_regression_data

This writes 4 files to data/phyloregression/.

Linear regression (brms)

pixi run -e regression beast_phylolm -- --model_type linear_geo --tree input_v12_combined_resampled --variant with_inventory
pixi run -e regression beast_phylolm -- --model_type linear_geo --tree heggarty2024_raw --variant with_inventory

GP regression (cmdstanr)

pixi run -e regression beast_phylolm -- --model_type gp_geo --tree input_v12_combined_resampled --variant with_inventory
pixi run -e regression beast_phylolm -- --model_type gp_geo --tree heggarty2024_raw --variant with_inventory

Results are written to data/phyloregression/<variant>/.

Plots

Install visualization dependencies:

pixi install -e viz

Publication figures

# Figure 1
pixi run -e viz fig1_acc_vs_brsupport  # Panel A: LID accuracy vs. bootstrap support
pixi run -e viz fig1_nmf               # Panel B: NMF structure plot
pixi run -e viz fig1_delta             # Panel D: per-language delta scores
pixi run -e viz fig1_pca               # Extended: PCA of XLS-R embeddings
pixi run -e viz fig1_sqa               # Extended: silhouette vs. SI-SDR + correlation

# Figures 2–3
pixi run -e viz fig2_rates             # Figure 2 panel B: speech rate over time
pixi run -e viz fig2_rates_cognate     # Cognate rate over time
pixi run -e viz fig3_geo              # Figure 3: regression panels

# Extended
pixi run -e viz ext_rates_and_maps     # rate scatter, GP maps, root age, rate-over-time

Compute paper stats

pixi run python -m src.tasks.phylo.compute_paper_stats