Non-negative Matrix Factorization with Selective Sparsity Constraints for Transcription of Bell Chiming Recordings

The paper presents a method for automatic transcription of recordings of bell chiming performances. Bell chiming is a Slovenian folk music tradition involving performers playing tunes on church bells by holding the clapper and striking the rim of a stationary bell. The tunes played consist of repeated rhythmic patterns into which various changes are included. Because the sounds of bells are inharmonic and their tuning not known in advance, we propose a two step approach to transcription. First, by analyzing the covariance matrix of the time-frequency representation of a recording, we estimate the number of bells and their approximate spectra using prior knowledge of church bell acoustics and bell chiming performance rules. We then propose a non-negative matrix factorization algorithm with selective sparsity constraints that learns the basis vectors that approximate the previously estimated bell spectra. The algorithm also adapts the number of basis vectors during learning. We show how to apply the proposed method to bell chiming transcription and present results on a set of field recordings.