Differences Between Singer and Speaker Verification: Training Singer Feature Representation Extractor Utilizing Singing Voice Characteristics

We aimed to construct a robust feature extractor for singer verification using a straightforward method that leverages the unique characteristics of singing voices. The speaker feature extractor based on ECAPA-TDNN needs to be trained to identify numerous speakers using voice data that vary within the same speaker and to distinguish voices that sound similar but belong to different speakers. However, collecting singing voice data is challenging, particularly when constructing a corpus with a large number of singing samples from the same singer. In this study, we address this problem by adopting a simple approach: segmenting singing voices before inputting them into ECAPA-TDNN during training. The validity of this approach arises not merely from increasing the amount of data, but from the inherent characteristics of singing voices. Specifically, compared to spoken voices, singing voices exhibit less acoustic variation over short periods but greater variation over long periods. By utilizing short segments of singing voices for training, we can develop a feature extractor that is robust to acoustic variations within the same singer. Our experiments demonstrate the effectiveness of the proposed approach of segmenting singing voices into three-second intervals for training and provide insights that this method does not yield the same benefits for spoken voices, highlighting its unique effectiveness for singer verification using singing voices.