mueller91/MLAAD

Introduction

Welcome to MLAAD: The Multi-Language Audio Anti-Spoofing Dataset -- a dataset to train, test and evaluate audio deepfake detection. See the paper for more information.

License:

Starting from MLAADv8, this dataset will be published under a non-commercial license (CC-BY-NC 4.0). If you want to use this dataset for commercial purposes, you need to either:

• use a previous version (MLAAD v1 - v7)
• contact us to obtain a commercial license ([email protected])

Download the dataset

Option 1: Hugging Face datasets library

Install the datasets package:

pip install datasets

Login with your Hugging Face account:

huggingface-cli login

Then load the dataset in Python:

from datasets import load_dataset

# Download from HF and cache
ds = load_dataset("mueller91/MLAAD")

# Optionally: Save the dataset to your own directory
ds.save_to_disk("MLAAD_local")

This will automatically handle authentication and download.

Option 2: Git + git-lfs

If you prefer to clone with git, you must first login via Hugging Face:

huggingface-cli login

Then clone:

git lfs install
git clone https://huggingface.co/datasets/mueller91/MLAAD

Structure

The dataset is based on the M-AILABS dataset. MLAAD is structured as follows:

fake
|-language_1
|-language_2
|- ....
|- language_K
    | - model_1_K
    | - model_2_K
    | - ....
    | - model_L_K
        | - meta.csv
        | - audio_L_K_1.wav
        | - audio_L_K_2.wav
        | - audio_L_K_3.wav
        | - ....
        | - audio_L_K_1000.wav

The file 'meta.csv' contains the following identifiers. For more in these, please see the paper and our website.

path|original_file|language|is_original_language|duration|training_data|model_name|architecture|transcript|reference_speaker

where reference_speaker may or not be present - this key has been introduced only in v8.

Proposed Usage

We suggest to use MLAAD either as new out-of-domain test data for existing anti-spoofing models, or as additional training resource. We urge to complement the fake audios in MLAAD with the corresponding authentic ones from M-AILABS, in order to obtain a balanced dataset. M-AILABS can be downloaded here. An antispoofing model trained on (among others) the MLAAD dataset is available here.

Bibtex

@article{muller2024mlaad,
  title={MLAAD: The Multi-Language Audio Anti-Spoofing Dataset},
  author={M{\"u}ller, Nicolas M and Kawa, Piotr and Choong, Wei Herng and Casanova, Edresson and G{\"o}lge, Eren and M{\"u}ller, Thorsten and Syga, Piotr and Sperl, Philip and B{\"o}ttinger, Konstantin},
  journal={arXiv preprint arXiv:2401.09512},
  year={2024}
}