FAD: A Chinese Dataset for Fake Audio Detection Under Noise

Twelve mainstream speech generation techniques were used to create fake audios for this dataset. It contains two versions, clean and noisy, with the noisy version created by adding noise from three databases at five different signal-to-noise ratios. The dataset is structured into training, development, and test sets, with a further split into seen and unseen subsets to evaluate model generalization.

Use Cases

• Fake audio detection based on audio samples generated by twelve different techniques.
• Algorithm source identification for audio forensics based on the generation method of fake utterances.
• Evaluating model robustness to additive noise based on the five SNR levels and three noise sources.
• Testing model generalization to unseen audio types based on the dedicated unseen test subset.

Strengths

• Designed to fill a gap as a standard public Chinese dataset for fake audio detection under additive noise conditions.
• Includes a controlled noisy version with five distinct SNR levels to simulate real-life scenarios.
• Features a structured evaluation split with seen and unseen test sets to specifically assess model generalization.

Limitations

• Key metadata such as row count, file size, specific column names, and last update date are not provided by any source.
• The description is repeated verbatim across multiple platform entries, suggesting potential metadata duplication rather than independent sourcing.

Provenance

Source

Haoxin Ma, Chinese Academy of Sciences

Collection Method

Fake audios were synthetically generated using twelve speech generation techniques; real audio source is not specified. Noises were added from three external noise databases.

Geography

China (Chinese language audio)