Hugging Face's logo

Datasets:
coml
/
vox-communis-artifacts

Tasks:
Automatic Speech Recognition
Modalities:
Text
Size:
1M - 10M
Tags:
phonetics
multilingual
License:
Dataset card Data Studio Files
xet
 Community
1
vox-communis-artifacts / README.md
angelo-ortiz's picture
angelo-ortiz
Upload folder using huggingface_hub
466c126 verified
preview code
|
Raw
History Blame Contribute Delete
7.45 kB
metadata 
license: cc0-1.0
pretty_name: VoxCommunis Artifacts
task_categories:
  - automatic-speech-recognition
tags:
  - phonetics
  - multilingual
source_datasets:
  - mozilla-foundation/common_voice_17_0
  - pacscilab/VoxCommunis
language_bcp47:
  - ab
  - am
  - ba
  - be
  - bg
  - bn
  - ca
  - ckb
  - cs
  - cv
  - dv
  - el
  - eu
  - fr
  - gn
  - ha
  - hi
  - hsb
  - hu
  - hy-AM
  - id
  - it
  - ja
  - ka
  - kk
  - kmr
  - ko
  - ky
  - lij
  - lt
  - ml
  - mn
  - mr
  - mt
  - myv
  - nan-tw
  - nl
  - pa-IN
  - pl
  - pt
  - ro
  - ru
  - rw
  - sk
  - sl
  - sq
  - sr
  - sv-SE
  - sw
  - ta
  - th
  - tk
  - tr
  - tt
  - ug
  - uk
  - ur
  - uz
  - vi
  - yo
  - yue
  - zh-CN
  - zh-HK

VoxCommunis Artifacts

The data artifacts used to train and evaluate the MauBERT models, introduced by the MauBERT paper: manifests, frame-level phone alignments, per-language phone inventories, and the language table.

Companion model repositories:

No audio is included

The recordings come from Common Voice 16.1, which you must download yourself; the phone annotations come from VoxCommunis. This repository only ships the derived annotations and the file lists that tie them to the Common Voice audio.

Language split

At the time of download (August 2024), there were 63 languages, out of which 5 languages were held-out as development languages (Swahili, Tamil, Thai, Turkish and Ukrainian) and 3 languages were unused due to their presence in ZRC2017, that is the test languages (French and two Chinese variants).

Files

  • vox_communis_languages.tsv: The 63 VoxCommunis languages, with family, branch, duration in minutes (before and after 50h capping), and a split column marking the 55 pre-training languages (train) against the 8 held-out languages used as development (dev) or test (test) languages.
  • inventories.jsonl: The IPA phone inventories, one language per line, {"language": <code>, "name": <name>, "phones": [...]}, exactly as they come out of the VoxCommunis alignments (after character normalisation).
  • canonical_inventories.jsonl: The same inventories, with every phone replaced by its PanPhon canonical representative (maubert.data.FeatureDecoder.segment_to_representative) — these are the segments the models actually index. This is the file the library reads: it is required to reduce the phone head when extracting features from the phone projection layer, for both model variants.
  • full_dataset/multilingual-{train,dev,test}.{tsv,align}: The original multilingual Common Voice splits (not used), plus the reduced subsets used for validation and testing of the multilingual pre-training, namely multilingual-dev-1h and multilingual-test-2h.
  • full_dataset/{train,dev,test,train-50h}/{manifests,alignments}/<code>.{tsv,align}: The same data as above, but separated by both split and language. train-50h is the duration-capped split actually used during multilingual pre-training.
  • dev_languages/{train-10min,train-1h,train-10h,dev,test}/{manifests,alignments}/: The extracted splits for the 5 held-out development languages used for zero- and few-shot acoustic unit discovey (see task 1 of ZeroSpeech challenge series), with the corresponding ABX items/ for the dev (released but never used due to the high number of speakers) and test sets.

Manifests (.tsv) list audio files: the first line is the corpus root, every subsequent line is <path relative to root>\t<number of samples>. As shipped, the root is the placeholder /path/to/cv-corpus-16.1 and every path is a bare filename; see Re-rooting the manifests to point them at your copy of the corpus. Alignments (.align) carry the frame-level phone labels, one line per utterance, <utterance id>\t<space-separated phones>.

Usage

Install maubert, then download this repository:

from huggingface_hub import snapshot_download

data_dir = snapshot_download("coml/vox-communis-artifacts", repo_type="dataset")

Re-rooting the manifests

Every manifest ships with the placeholder root /path/to/cv-corpus-16.1 on its first line, and its rows are bare .wav filenames. Point the manifests at your own copy of the corpus before training:

from maubert.data.io import update_manifest

update_manifest(
    f"{data_dir}/full_dataset/multilingual-train.tsv",
    "/data/cv-corpus-16.1",
    file_extension=".mp3",
)

Matching is done on the file stem, which is identical between the original .mp3 and the converted .wav, so this works against an unconverted Common Voice release and against any directory layout under the root you pass — only pass the extension your copy actually uses. However, beware that MauBERT models were trained on and accept .wav files only.

The helper rewrites the manifest in place and leaves the previous version alongside it as multilingual-train.tsv.bak; it raises if that backup already exists, so run it once per download. It also raises if any utterance in the manifest is missing from your tree.

Once re-rooted, the files map directly onto the data section of configs/train_feat.yaml and configs/train_phone.yaml in the maubert repository — full_dataset/train-50h/{manifests,alignments} as train_manifest / train_alignment and full_dataset/multilingual-dev-1h.{tsv,align} as val_manifest / val_alignment, while canonical_inventories.jsonl fills the downstream.inventory_file field used by the monolingual configs.

Looking up a language inventory 

from maubert.data.phone_dataset import get_language_inventory

# Either the language name or its BCP-47 code
inventory = get_language_inventory(f"{data_dir}/canonical_inventories.jsonl", "italian")

Citing

Please cite MauBERT alongside Common Voice and VoxCommunis:

@inproceedings{ortiztandazo-etal-2026-maubert,
    title = "{M}au{BERT}: Universal Phonetic Inductive Biases for Few-Shot Acoustic Units Discovery",
    author = "Ortiz Tandazo, Angelo  and
      Khentout, Manel  and
      Benchekroun, Youssef  and
      Hueber, Thomas  and
      Dupoux, Emmanuel",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.24/",
    doi = "10.18653/v1/2026.acl-long.24",
    pages = "568--585",
    ISBN = "979-8-89176-390-6",
}