ATRemix: An Auto-tune Remix Dataset for Singer Recognition

Lifang Wang, Bingyuan Wang, Guanghao Tan, Wei-Qiang Zhang, Jun Feng, Bing Zhu, Shenjin Wang

Abstract: In recent years, with the development of video websites such as YouTube, TikTok, and Bilibili, a great number of auto-tune remix audios are produced every day. Auto-tune remix audios are usually made from existing famous audios. The original clips can be tuned to various remixes through professional editing techniques. In the creation process, the characteristics of the singer are usually maintained, thus the original materials can be traced by singer recognition methods. This paper mainly focuses on the research of auto-tune remix singer recognition. As this topic has not been discussed before, we create a dataset of auto-tune remix audios and attempt to recognize the identity of the singer. Firstly, we use an x-vector model trained on the TIMIT dataset, and then evaluate it on the ATRemix dataset. Secondly, ATRemix dataset used to train different models, and SubATRemix dataset used as a test set, which shows good performance.