This paper was accepted on the IEEE Workshop on Purposes of Sign Processing to Audio and Acoustics (WASPAA) 2025
Non-negative Matrix Factorization (NMF) is a strong method for analyzing regularly-sampled information, i.e., information that may be saved in a matrix. For audio, this has led to quite a few purposes utilizing time-frequency (TF) representations just like the Quick-Time Fourier Rework. Nevertheless extending these purposes to irregularly-spaced TF representations, just like the Fixed-Q rework, wavelets, or sinusoidal evaluation fashions, has not been doable since these representations can’t be instantly saved in matrix kind. On this paper, we formulate NMF by way of learnable capabilities (as a substitute of vectors) and present that NMF may be prolonged to a greater diversity of sign lessons that needn’t be frequently sampled.
- † College of Illinois at Urbana-Champaign
