ResMamba: A Dual-Path Convolution-Mamba Model for Music Genre Classification
Abstract
Music genre classification remains a fundamental yet challenging task in music information retrieval, requiring models that capture local time--frequency patterns and long-range temporal dependencies. To address this, we propose ResMamba, a hybrid architecture that augments a ResNet-style four-stage backbone with Mamba blocks. Building upon the selective structured state space sequence model, ResMamba first adapts the selective SSM into a 4D-compatible residual Mamba block that captures global dependencies at linear complexity. We then design the ConvMamba dual-path block, which combines this adapted Mamba branch with a parallel convolutional branch through a learned fusion, jointly modeling global sequential context and local spectro-temporal patterns. Furthermore, we adopt a multi-scale fusion head that aggregates hierarchical representations across all four stages. Experiments on the GTZAN dataset demonstrate that our full ResMamba model achieves a 5-fold cross-validation accuracy of , surpassing the strongest published method in our comparison (88.64%). All models are trained from scratch without any pretraining, external data, or data augmentation. Beyond GTZAN, ResMamba also generalizes to the larger FMA benchmarks under the same training protocol, attaining on FMA-Small under 10-fold cross-validation --- the best published from-scratch result --- and on FMA-Medium under the official artist-level split, surpassing the widely cited SVM baseline of Defferrard et al. (63%).
Figures & tables
| Category | Parameter | Value |
| Audio preprocessing | Sample rate | 22050 Hz |
| 2048 | ||
| Hop length | 512 | |
| 128 | ||
| Segment duration | 30 s | |
| Mel-spectrogram size |
| Resolution | Comp. | Feature Map | Acc (%) |
|---|---|---|---|
| Model | Block Type | nBlocks | Acc. (%) | Params (M) |
|---|---|---|---|---|
| ResNet-18 | BasicBlock | 11.2 | ||
| ResMamba | AdaptMamba | 2.2 | ||
| ResMamba | ConvMamba | 14.6 | ||
| ResMamba | ConvMamba+MS | 15.1 | ||
| ResMamba | ConvMamba+MS | 28.0 |
| Model | Acc (%) | Params (M) | FLOPs (G) |
|---|---|---|---|
| ResNet-18 | 11.2 | 12.1 | |
| ResNet-34 | 21.3 | 24.3 | |
| ResNet-50 | 23.5 | 27.2 | |
| ResNet-101 | 42.5 | 51.8 | |
| EfficientNet-B0 | 4.0 | 2.6 | |
| EfficientNet-B1 | 6.5 | 3.9 |
| Method | Acc (%) | Split |
| Chen et al. [ 23 ] | 87.41 | 10-fold, segment-level a |
| Gan et al. [ 34 ] | 85.5 | 80/20 train/test b |
| Wu et al. [ 31 ] | 86.8 | 80/20 train/val b |
| Setiorini et al. [ 33 ] | 88.64 | 60/20/20, segment-level a |
| ResMamba (5-fold CV) | 5-fold CV c |
| Method | Acc. (%) | Split |
| Heakl et al. [ 19 ] | 58.30 | 10-fold CV a |
| Muniz et al. [ 26 ] | 60.30 | Nested 10-fold CV b |
| Seo et al. [ 29 ] | 57.38 0.94 | 80/20 10 c |
| ResNet-18 | 59.95 1.81 | 10-fold CV d |
| ResMamba | 64.52 1.88 | 10-fold CV e |
| Method | Acc. (%) | Macro F1 | Split |
|---|---|---|---|
| Defferrard et al. [ 17 ] | 63.00 | — | Official artist split |
| ResNet-18 † | 62.99 | 0.3558 | Official artist split |
| ResNeXt-50 † (32 4d) | 61.00 | 0.3516 | Official artist split |
| ResMamba † | 65.75 | 0.4262 | Official artist split |
Music Information Retrieval
Music Information Retrieval develops computational methods for finding, classifying, describing, and analyzing music from audio, scores, and metadata.
See all 165 papers