Vocal Singing and Music Separation of Mizo Folk Songs

Music source separation is an active research topic in the area of music processing. With the recent advancements of unsupervised learning models, there is an increase in the influx of research with increased performances and improved results over existing traditional methods. In this work, we show how music source separation can be used for low-resource Mizo folk songs. We employed different methods of separating background music and vocals, as well as different sources in the background music. These separations are carried out using existing source separation algorithms based on repeating patterns, matrix decomposition, and PCA, namely REpeating Pattern Extraction Technique (REPET), two variants of Non-negative Matrix Factorization (NMF), and Robust Principal Component Analysis (RPCA). A Mizo folk song dataset, consisting of 60 songs from three categories of Mizo folk songs- Pawnto Hta (children’s songs), Dâr Hta (Gong songs), and Bawh hta and Hlado (War and Hunting chants), is used. The employed methods perform well and achieve good SDR and SNR values, with RPCA achieving the best values.