Musical instrument identification using multiscale Mel-frequency cepstral coefficients

We investigate the benefits of evaluating Mel-frequency cepstral coefficients (MFCCs) over several time scales in the context of automatic musical instrument identification for signals that are monophonic but derived from real musical settings. We define several sets of features derived from MFCCs computed using multiple time resolutions, and compare their performance against other features that are computed using a single time resolution, such as MFCCs, and derivatives of MFCCs. We find that in each task - pair-wise discrimination, and one vs. all classification - the features involving multiscale decompositions perform significantly better than features computed using a single time-resolution.