EmoFake / README.md
01Yassine's picture
Publish verified audio data and metadata
d2ccb37 verified
|
Raw History Blame Contribute Delete
3.95 kB
metadata
license: cc-by-nc-nd-4.0
language:
  - en
  - zh
task_categories:
  - audio-classification
tags:
  - speech-deepfake-detection
  - emotional-speech
  - audio
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: audio/train-*.parquet
      - split: validation
        path: audio/val-*.parquet
      - split: test
        path: audio/test-*.parquet

EmoFake — DeepFense data release

This repository redistributes the EmoFake audio and official protocol partitions used by DeepFense. It credits the original dataset authors; DeepFense is not the creator of the source recordings or spoof systems.

Source: https://zenodo.org/records/12806228

Attribution: Yan Zhao, Jiangyan Yi, Jianhua Tao, Chenglong Wang and Yongfeng Dong. EmoFake: An Initial Dataset for Emotion Fake Audio Detection. CCL 2024, pp. 1286–1297. https://aclanthology.org/2024.ccl-1.99/

License: cc-by-nc-nd-4.0. The source terms remain in force; no commercial-use rights or new license are granted by this packaging. All audio bytes are retained unchanged. There is no resampling, trimming, normalization, denoising or lossy conversion in this release.

Official splits

Source split Rows
train 54,600
val 18,200
test 35,000

Total: 107,800 recordings. Source dev is named val in the original SDD metadata and validation in the default Hugging Face dataset configuration. Labels are the strings bonafide (genuine) and spoof (manipulated). These are the supplied partitions; no random split was invented.

Load portable audio

from datasets import load_dataset

ds = load_dataset("DeepFense/EmoFake", split="train", streaming=True)

The default dataset reads audio/*.parquet. Each row contains an audio feature with the original WAV bytes, original protocol metadata, audio_sha256, a logical repository-relative path, and the original absolute sdd_path for provenance. The audio feature loads from embedded bytes; it does not require /ds-slt or any cluster directory. Audio decoding uses the audio backend supported by your installed Datasets version.

To access the exact file bytes without decoding or resampling:

from datasets import Audio

raw = ds.cast_column("audio", Audio(decode=False))
row = next(iter(raw))
payload = row["audio"]["bytes"]

path is the preserved logical WAV filename, rather than a separate loose file in this repository. To recreate the local audio tree, write payload to that relative path. audio_sha256 verifies it against the original file. The exporter verifies every embedded payload after writing each shard.

Original SDD Parquets and protocols

  • data.parquet: original full SDD table, copied byte-for-byte.
  • parquets/train.parquet, parquets/val.parquet, parquets/test.parquet: original SDD protocol files, copied byte-for-byte.
  • protocols/: original source label files.
  • SOURCE_SDD_README.md: local import documentation, retained as provenance.
  • publication.json, SHA256SUMS: export checks, exact counts and checksums.

The original four metadata Parquets keep their original absolute path and original_path columns. They are preserved for reproducibility and are not the portable audio loader. To train existing path-based scripts on a different machine, extract the embedded bytes under their logical relative paths and rebase a local copy of the metadata. Do not point a new machine at the original cluster paths.

Scope

This release contains the complete imported audio and all supplied train/validation/test rows. Protocol labels and source metadata are preserved. Dataset-specific metadata and any source conditions are described in the retained source documentation. The publisher changes storage format only; it does not change the waveform bytes or claim source-dataset authorship.