HR-Mel: An Optimized Asymmetric Multi-Band Mel Representation for Music Signals

We present HR-Mel, an optimized fixed 3-band, 96-bin mel representation for 44.1 kHz music analysis with asymmetric bin allocation (48/32/16 bins for 0-1.2 kHz, 1.2-5 kHz, and 5-20 kHz bands) and band-specific pointwise nonlinearities (log1p / log1p / sqrt(log1p)). The configuration is selected by a design-space search on a 12-track exploratory set and evaluated without further tuning on the GTZAN benchmark (999 tracks). On GTZAN, HR-Mel reduces pseudo-inverse STFT reconstruction error from 0.3747 to 0.1730 relative to a same-dimensionality 96-bin log-mel baseline, corresponding to a 53.8% improvement. Under matched nominal-rate scalar quantization, HR-Mel maintains a large rate-distortion advantage (e.g., 8-bit: 0.1733 vs. 0.3748). A downstream genre classification experiment shows comparable or slightly improved accuracy (66.0% vs. 64.7%). This is the optimized version (v2) following a design-space search, superseding the earlier exploratory prototype (v1, 40/32/24 allocation). Code: https://github.com/mini-kio/HR-Mel