Auto-Tune Detection Dataset (ATD-Dataset)

We present a novel dataset for distinguishing auto-tuned musical compositions from authentic ones, filling a gap in existing resources. To create this dataset, we utilized the VocalSet and Musdb18 datasets from the Music Information Retrieval (MIR) domain. The dataset comprises 150 music recordings along with 3613 audio recordings from 20 professional singers. It is segmented into 'Training' and 'Test' subsets, with 50 music recordings allocated to the test subset and the remainder to the training subset. The following explains the file names and subdirectory names:

·        Simple: This directory contains song tracks in the test dataset without any postprocessing.

·        MP3_Compression: This directory contains MP3-compressed songs in the test dataset.

·        Random_Aug: This directory contains songs in the test dataset after random augmentation.

·        Auto_Tuned.wav: Auto-tuned versions of the vocal performances.

·        Original.wav: Original vocal performances.

·        st_Auto_Tuned.wav: Auto-tuned versions of the vocal performances, standardized to be a 10-second audio file.

·        st_Original.wav: Original vocal performances, standardized to be a 10-second audio file.

·        Auto_Tuned_Vocal.wav: Auto-tuned vocals of songs.

·        Original_Vocal.wav: Original vocals of songs.

·        Auto_Tuned_Song.wav: Auto-tuned songs.

·        Original_Song.wav: Original songs.

·        Auto_Tuned_Vocal_is.wav: Auto-tuned vocals extracted from mixed songs using the vocal isolator tool.

·        Original_Vocal_is.wav: Original vocals extracted from mixed songs using the vocal isolator tool.