AI xPress

AI and computer science research.

Back to Foragentis

Speech & Audio

A 23-language audio dataset outperformed comparable training datasets in voice-deepfake detection tests

Speech & Audio November 9, 2026

Headphones on sheet music on a piano

In a 2024 paper, researchers report that a new dataset of synthetic voices in 23 languages led to better detection of fabricated voice recordings when used to train three deepfake detection models. The dataset, called MLAAD, contains 160.2 hours of synthetic speech made with 52 text-to-speech models across 22 architectures. The work appeared in the International Joint Conference on Neural Networks (IJCNN).

What problem does the dataset address?

Text-to-speech tools can give a voice to people with speech impairments, but they can also create fake voices that mislead listeners and defeat voice-based security checks. The researchers point out that existing anti-spoofing data concentrate heavily on English and Chinese audio, which limits how well detection models work in other languages.

What did the researchers test?

The team used MLAAD to train three deepfake detection models. The paper reports that MLAAD produced stronger results than two comparable datasets, InTheWild and FakeOrReal, when used as a training resource.

How did MLAAD compare with ASVspoof 2019?

Across eight benchmark datasets, MLAAD and the established ASVspoof 2019 dataset each led on four. The authors describe MLAAD as a complementary resource to ASVspoof 2019 rather than a replacement.

FAQ

What is MLAAD?

MLAAD stands for Multi-Language Audio Anti-Spoofing Dataset. It contains 160.2 hours of synthetic voice in 23 languages, made with 52 text-to-speech models that use 22 different architectures.

Did the researchers release the dataset?

The abstract says the authors published the dataset and made trained models available through an interactive web server.

Does the study show real-world detection?

No. The results come from benchmark tests, not from a field trial.

Limits and what's next

The evaluation is a benchmark study. The researchers trained three detection models on MLAAD and compared them on benchmark datasets, including InTheWild, FakeOrReal and ASVspoof 2019. In one comparison across eight datasets, MLAAD and ASVspoof 2019 each led on four. The abstract reports no independent replication of the result and no separate field test. The authors state a goal of making anti-spoofing tools available beyond specialists, but the abstract does not describe testing those tools in live conditions.

Source: Nicolas M. Müller et al., “MLAAD: The Multi-Language Audio Anti-Spoofing Dataset,” International Joint Conference on Neural Networks (IJCNN), 2024.