Table 3 LSD performance of the evaluated models for monophonic and polyphonic real-world datasets. Best models are shown in bold for each dataset. The last two columns show CPU inference time expressed as real-time percentage and the number of parameters of the models
From: Efficient bandwidth extension of musical signals using a differentiable harmonic plus noise model
Model | Log-spectral distance | Inference time | # of parameters | |||
|---|---|---|---|---|---|---|
Monophonic datasets | Polyphonic datasets | |||||
OrchideaSOL | Medley-solos-db | MedleyDB | Gtzan | (% real-time) | ||
Null | 15.9 | 18.53 | 24.37 | 33.84 | 0 | 0 |
SBR [13] | 9.27 | 8.78 | 11.15 | 12.96 | 2 | 0 |
ResNet [12] | 14.04 | 15.65 | 16.17 | 26.84 | 48 | 55M |
DDSP-noise | 7.20 | 8.28 | 8.96 | 10.06 | 3 | 3.5M |
DDSP-mono-dec | 5.68 | 8.09 | 8.98 | 9.95 | 9 | 4.4M |
DDSP-mono-dec-cyclic | / | / | 11.57 | 11.60 | 44 | 4.4M |
DDSP-poly-dec | / | / | 9.53 | 10.31 | 9 | 7.5M |