PartialEdit: Identifying Partial Deepfakes in the Era of Neural Speech Editing

This is part of the dataset we curated based on VCTK to study partial speech deepfake detection in the era of neural speech editing. For more details, please refer to our Interspeech 2025 paper: "PartialEdit: Identifying Partial Deepfakes in the Era of Neural Speech Editing". In the paper, we curated four subsets: E1: VoiceCraft, E2: SSR-Speech, E3: Audiobox-Speech, and E4: Audiobox. Adhering to Audiobox's license, we cannot release the E3 and E4 subsets. V1.1 updates from V1: We corrected a few errors of the timestamps and added the speaker splits. The folder structure is as follows: PartialEdit/├── PartialEdit_E1E2.csv├── E1/│ ├── p225/│ │ ├── p225_001_edited_partial_16k.wav│ │ ├── p225_002_edited_partial_16k.wav│ │ └── ...│ ├── p231/│ │ ├── p231_001_edited_partial_16k.wav│ │ ├── p231_002_edited_partial_16k.wav│ │ └── ...│ └── ...├── E1-Codec/│ └── (same structure as E1)├── E2/│ └── (same structure as E1)├── E2-Codec/│ └── (same structure as E1)└── modified_txt/├── p225/│ ├── p225_001_modified.txt│ ├── p225_002_modified.txt│ ├── p225_003_modified.txt│ └── ...├── p231/│ ├── p231_001_modified.txt│ ├── p231_002_modified.txt│ └── ...└── ... Paper: https://www.isca-archive.org/interspeech_2025/zhang25g_interspeech.html Demo page: https://yzyouzhang.com/PartialEdit/index.html The `PartialEdit_E1E2.csv` file contains information about the edited regions in each audio file. Each row represents the following columns: - `filename`: The name of the audio file.- `start of the edited region (s)`: The starting time (in seconds) of the first edited region.- `end of the edited region (s)`: The ending time (in seconds) of the first edited region.- `total duration (s)`: The total duration (in seconds) of the audio file. If there are two edited regions within a file, the row format expands to include: - `filename`: The name of the audio file.- `start of the edited region (s)`: The starting time (in seconds) of the first edited region.- `end of the edited region (s)`: The ending time (in seconds) of the first edited region.- `start of the second edited region (s)`: The starting time (in seconds) of the second edited region.- `end of the second edited region (s)`: The ending time (in seconds) of the second edited region.- `total duration (s)`: The total duration (in seconds) of the audio file. To make sure the download is complete, you can check the MD5 code with the following command: md5sum *