Half-Truth: A Partially Fake Audio Detection Dataset (HAD)

Several promising datasets have been developed to advance the field of fake audio detection. However, these previous datasets have failed to address a critical scenario: the presence of an attacker who covertly inserts small fabricated audio clips into authentic speech recordings. This situation poses a significant security threat because differentiating these small fake segments from the overall speech utterance is an exceptionally challenging task. In response to this challenge, we introduce a groundbreaking dataset designed for the detection of partial audio falsifications, which we term Half-Truth Audio Detection (HAD). The partially manipulated audio samples contained within the HAD dataset involve minimal alterations, typically limited to modifying a few words within an utterance. These altered audio segments are created using state-of-the-art speech synthesis technology.This dataset not only empowers the identification of counterfeit utterances but also enables the pinpointing of manipulated regions within a speech recording.

When you use this dataset, please cite us:

Jiangyan Yi, Ye Bai, Jianhua Tao, Haoxin Ma, Zhengkun Tian, Chenglong Wang, Tao Wang, Ruibo Fu:Half-Truth: A Partially Fake Audio Detection Dataset. Interspeech 2021: 1654-1658