Effects of Time and Frequency Resolution on Nonnegative Matrix Factorization

In this paper, we present the comparison of note events extraction using nonnegative matrix factorization (NMF) with input from a classic Fourier transform (FT) and with input from an improved time resolution FT. In order to improve time resolution, FT sample window length must be increased. As the window length increases, the time resolution is sacrificed for a better frequency resolution. Hence, good time and frequency resolutions are required in our polyphonic music transcription task. Here, we first apply zero-padding algorithm to classic FT to help maintain the time and frequency resolution. Then, we apply the hanning window. Finally we use NMF to decompose the note events. The experiment shows that NMF performs the decomposition of multivariate nonnegative data matrix well after applying the improved time resolution FT.