Evaluation of spectral transforms for music signal analysis

In this paper we present a study on the spectral analysis of music signals comparing the time domain representation, the short-time Fourier transform (STFT) and the constant-Q transform (CQT) which are additionally combined with different signal-dependent transforms. The comparison is carried out with respect to the spectral compactness, the data compression ability and the temporal continuity of transform coefficients for which we propose measures in this paper. In addition, we investigate the performance of these transforms in a source separation task in which we strive for recovering the main melody line from a mixed instrument recording. Our experiments reveal that performing a rank-reduced principal component analysis based on a CQT representation exhibits the best results in terms of instrumental source separation measures and listening impression which points towards the potential of the CQT for improving existing source separation methods which are currently often based on the STFT.