Spotting Parodies: Detecting Alignment Collapse Between Lyrics and Singing Voice
Tomoki Ariga, Yosuke Higuchi, Mitsunori Kanno, Rie Shigyo, Takato Mizuguchi, Naoki Okamoto, Tetsuji Ogawa · 2023
We present a method for detecting parodies in karaoke singing by evaluating alignment collapse between lyrics and singing voice. Parody detection is a crucial technique for catching copyright infringement in online vocal recordings and enhancing the accuracy of singing scoring algorithms. Assuming ground-truth transcriptions (i.e., lyrics) are fixed for a specific song, we can use an acoustic model to obtain a forced alignment between lyrics and the corresponding singing voice. However, if the lyrics are not sung correctly and contain parodies, the alignment may not be estimated accurately due to mismatches in acoustic features. Our aim is to utilize this alignment collapse to create a robust and effective system for detecting parodies in singing voice. To this end, we explore various metrics to assess the degree of collapse in the estimated alignments. We also design metrics robust against singers' variations, such as differences in rhythm arrangements. In our experiments, we construct a Japanese singing voice dataset for training an acoustic model specific to the singing domain, which is essential for estimating proper alignments. The results demonstrate that the proposed system, adopting an edit-distance-based metric, is highly effective in detecting parodies, achieving the area under the curve (AUC) of 0.963.