HQ-MPSD: A Multilingual Benchmark for Partial Deepfake Speech Detection

HQ-MPSD Dataset v1

The structure of dataset folder is below:

HQ-MPSD/
├── English/
│ ├── Bonafide/
│ │ ├──<speakerID>_<audiobookId>_<segmentID>.flac
│ │ └── ...
│ ├── Fully_Fake/
│ │ ├── <speakerID>_<audiobookId>_<segmentID>_f.flac
│ │ └── ...
│ ├── Partial_Fake_Clean/
│ │ ├── <speakerID>_<audiobookId>_<segmentID>_p.flac
│ │ └── ...
│ └── ...
│ ├── Partial_Fake_Noisy/
│ │ ├── <speakerID>_<audiobookId>_<segmentID>_pn.flac
│ │ └── ...
│ └── Frame_level_label.txt
│ └── Noise_augmentation_info.txt
├── French/
│ └── (same structure as English)
├── German/
│ └── (same structure as English)
└── ...

Frame labels are defined as:

• 0 → Bonafide frame

• 1 → Deepfake frame

• 2 → Transition frame

Each row in the Frame_level_label.txt is formatted as follows: 
<id> <label_for_each_30ms_frame>

This file provides detailed provenance for noise-augmented audio samples, including:

• the augmentation type or noise category

• the specific noise file from OpenSLR26

• the specific noise file from MUSAN

Each row in the Noise_augmentation_info.txt is formatted as follows:
<id> <augmentation_label> <used_file_path_in_OpneSLR26> <<used_file_path_in_MUSAN>