XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark

by Ioan-Paul Ciobanu, Andrei-Iulian Hiji, Nicolae-Catalin Ristea, Paul Irofti, Cristian Rusu, Radu Tudor Ionescu

---

License

The source code and models are released under the Creative Common Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.

Reference

If you use this dataset or code in your research, please cite the corresponding paper:

1. Ioan-Paul Ciobanu, Andrei-Iulian Hiji, Nicolae-Catalin Ristea, Paul Irofti, Cristian Rusu, Radu Tudor Ionescu. XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark. arXiv preprint arXiv:2506.00462 (2025).

Bibtex:

@article{Ciobanu2025xmad,
  title="{XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark}",
  author={Ciobanu, Ioan-Paul and Hiji, Andrei-Iulian and Ristea, Nicolae-Catalin and Irofti, Paul and Rusu, Cristian and Ionescu, Radu Tudor},
  journal={arXiv preprint arXiv:2506.00462},
  year={2025}
}

Description

Recent advances in audio generation led to an increasing number of deepfakes, making the general public more vulnerable to financial scams, identity theft, and misinformation. Audio deepfake detectors promise to alleviate this issue, with many recent studies reporting accuracy rates close to $99%$. However, these methods are typically tested in an in-domain setup, where the deepfake samples from the training and test sets are produced by the same generative models. To this end, we introduce XMAD-Bench, a large-scale cross-domain multilingual audio deepfake benchmark comprising 668.8 hours of real and deepfake speech. In our novel dataset, the speakers, the generative methods, and the real audio sources are distinct across training and test splits. This leads to a challenging cross-domain evaluation setup, where audio deepfake detectors can be tested in the wild. Our in-domain and cross-domain experiments indicate a clear disparity between the in-domain performance of deepfake detectors, which is usually as high as $100%$, and the cross-domain performance of the same models, which is sometimes similar to random chance. Our benchmark highlights the need for the development of robust audio deepfake detectors, which maintain their generalization capacity across different languages, speakers, generative methods, and data sources.

Split statistics on our data set:

Results obtained with various state-of-the-art methods on our data set:

Download data

Our data is available at: https://drive.google.com/drive/folders/1PjboiIGjNWU6UeuIHrZu3ofF70o0A5-X?usp=drive_link

Detection framework

Modify the detection/config.json with the desired locations. Then run:

python detection/main.py

Demo generation script

python demo_script.py \
  --sentence "Generarea unui exemplu de test a reusit." \
  --refs ref1.wav ref2.wav ref3.wav ... \
  --output synthesized_sample.wav

Romanian datasets generation

VITS + FreeVC

    # in vits_freevc.py you need to modify the model to:
    vc_freevc = TTS("voice_conversion_models/multilingual/vctk/freevc24")
    model_name = "tts_models/ro/cv/vits"

VITS + KNN-VC

    # in vits_knnvc.py you need to modify the model to:
    knn_vc = torch.hub.load('bshall/knn-vc', 'knn_vc', prematched=True, trust_repo=True, pretrained=True)
    model_name = "tts_models/ro/cv/vits"

VITS + OpenVoice

    # in vits_openvoice.py you need to modify the model to:
    vc_openvoice = TTS("voice_conversion_models/multilingual/multi-dataset/openvoice_v2")
    model_name = "tts_models/ro/cv/vits"

Arabic datasets generation

fairseq + FreeVC

    # in fairseq_freevc.py you need to modify the model to:
    vc_freevc = TTS("voice_conversion_models/multilingual/vctk/freevc24")
    model_name = "tts_models/ara/fairseq/vits"

fairseq + KNN-VC

    # in fairseq_knnvc.py you need to modify the model to:
    knn_vc = torch.hub.load('bshall/knn-vc', 'knn_vc', prematched=True, trust_repo=True, pretrained=True)
    model_name = "tts_models/ara/fairseq/vits"

XTTSv2

    # in xttsv2.py you need to modify the model to:
    language = "ar"

Russian datasets generation

VITS + KNN-VC

    # in vits_knnvc.py you need to modify the model to:
    knn_vc = torch.hub.load('bshall/knn-vc', 'knn_vc', prematched=True, trust_repo=True, pretrained=True)
    model_name = "tts_models/rus/fairseq/vits"

VITS + OpenVoice

    # in vits_openvoice.py you need to modify the model to:
    vc_openvoice = TTS("voice_conversion_models/multilingual/multi-dataset/openvoice_v2")
    model_name = "tts_models/rus/fairseq/vits"

XTTSv2

    # in xttsv2.py you need to modify the model to:
    language = "ru"

English datasets generation

VITS + KNN-VC

    # in vits_knnvc.py you need to modify the model to:
    knn_vc = torch.hub.load('bshall/knn-vc', 'knn_vc', prematched=True, trust_repo=True, pretrained=True)
    model_name = "tts_models/en/ljspeech/vits"

VITS + OpenVoice

    # in vits_openvoice.py you need to modify the model to:
    vc_openvoice = TTS("voice_conversion_models/multilingual/multi-dataset/openvoice_v2")
    model_name = "tts_models/eng/fairseq/vits"

XTTSv2

    # in xttsv2.py you need to modify the model to:
    language = "en"

German datasets generation

VITS + KNN-VC

    # in vits_knnvc.py you need to modify the model to:
    knn_vc = torch.hub.load('voice_conversion_models/multilingual/multi-dataset/knnvc', 'knn_vc', prematched=True, trust_repo=True, pretrained=True)
    model_name = "tts_models/de/css10/vits-neon"

XTTSv2

    # in xttsv2.py you need to modify the model to:
    language = "de"

Spanish datasets generation

VITS + OpenVoice

    # in vits_openvoice.py you need to modify the model to:
    vc_openvoice = TTS("voice_conversion_models/multilingual/multi-dataset/openvoice_v2")
    model_name = "tts_models/spa/fairseq/vits"

XTTSv2

    # in xttsv2.py you need to modify the model to:
    language = "es"

Mandarin datasets generation

Tacotron + KNNVC

    # in vits_knnvc.py you need to modify the model to:
    knn_vc = torch.hub.load('bshall/knn-vc', 'knn_vc', prematched=True, trust_repo=True, pretrained=True)
    model_name = "tts_models/zh-CN/baker/tacotron2-DDC-GST"

Bark + FreeVC

    # in vits_freevc.py you need to modify the model to:
    vc_freevc = TTS("voice_conversion_models/multilingual/vctk/freevc24")
    model_name = "tts_models/multilingual/multi-dataset/bark"