Steven Kiemyang Tjoa (University of Maryland), “Sparse and nonnegative factorizations for music understanding” (2011)

Steven Kiemyang Tjoa (University of Maryland), “Sparse and nonnegative factorizations for music understanding” (2011), Advisor: Prof. K. J. Ray Liu

In this dissertation, the author proposes methods for sparse and nonnegative factorization that are specifically suited for analyzing musical signals. First, two constraints that aid factorization of musical signals: harmonic and co-occurrence constraints are discussed. Using the proposed constraints, when there is significant spectral-temporal overlap among the musical sources, the proposed method outperforms popular existing matrix factorization methods as measured by the recall and precision of learned dictionary atoms. Furthermore, the ability of representing each musical note with multiple atoms and clustering the atoms for source separation purposes is demonstrated. Second, spectral and temporal information extracted by nonnegative factorizations is used to improve musical instrument recognition. Third, the author studies how to perform sparse factorization when a large dictionary of musical atoms is already known.

For details, please access the full thesis here.