Score-informed source separation of choral music¶
Choral music is a challenging target for source separation due to the choral blend and the inherent acoustical complexity of the ‘choral timbre’. We compare two methods: score-informed NMF (non-negative matrix factorization, an unsupervised method) and Wave-U-Net (a convolutional neural network). To tackle the lack of datasets for training we create a synthesized choral dataset. We extend Wave-U-Net to condition the separation process on the musical score, and show that using the score improves separation performance, especially for the inner voices.
Matan Gover is a multi-disciplinary musician and software developer. He currently works as a researcher on singing voice technologies at Voctro Labs (Barcelona) focusing on singing voice synthesis. Matan completed his B.Mus. in orchestral conducting at the Jerusalem Academy of Music, and his M.A. in Music Technology at McGill University (Montreal), where he focused on computational processing of vocal music. Matan is also a professional pianist and choir singer, and has written orchestrations and arrangements for various vocal and instrumental ensembles.