Evaluation of mel-band and MFCC-based error metrics for correspondence to discrimination of spectrally altered musical instrument sounds

Several mel-band-based metrics and a single MFCC-based error metric were evaluated for best correspondence with human discrimination of single tones resynthesized from similar musical instrument time-varying spectra. Results show high levels of correspondence that are very close and often nearly identical to those found previously for harmonic and critical-band error metrics. The number of spectrum-related terms in the metrics required to achieve 85% R² correspondence is about five for harmonics, ten for mel bands, and ten for MFCCs, leading to the conjecture that subjects discriminate more on the basis of the first few harmonics than on the broad spectral envelope.