Source-linked AI summary
Bottom-up Broadcast Neural Network For Music Genre Classification
Caifeng Liu, Lin Feng, Guochao Liu, Huibing Wang, Shenglan Liu
TL;DR
CNN music classifiers often reuse image-recognition structures that may not suit spectrograms or preserve useful low-level features. This paper proposes BBNN, which combines multi-scale Inception blocks and dense connections to carry spectrogram information toward genre decisions. The authors report comparisons with state-of-the-art methods across benchmark datasets, including 97.2% accuracy on Extended Ballroom.
Problem
CNN-based music classifiers often apply image-recognition architectures to spectrograms without modification, while genre-relevant time-frequency scales and low-level features may require different handling.
Source: §1 · p. 2Method
BBNN combines multi-scale Inception blocks with dense connectivity to preserve and transmit low-level spectrogram features toward the decision layers.
Source: §1 · p. 2, §2.1 · p. 3Results
97.2% accuracy on Extended Ballroom; BBNN surpasses the compared state-of-the-art CNN-based and hand-crafted-feature methods on that dataset.
Source: §3.6 · p. 6Takeaways & Limitations
The authors report that BBNN compares favorably with state-of-the-art methods across benchmark datasets, including GTZAN, Ballroom, and Extended Ballroom.
Source: §4 · p. 6
Abstract
from arXiv · showhide
Music genre recognition based on visual representation has been successfully explored over the last years. Recently, there has been increasing interest in attempting convolutional neural networks (CNNs) to achieve the task. However, most of existing methods employ the mature CNN structures proposed in image recognition without any modification, which results in the learning features that are not adequate for music genre classification. Faced with the challenge of this issue, we fully exploit the low-level information from spectrograms of audios and develop a novel CNN architecture in this paper. The proposed CNN architecture takes the long contextual information into considerations, which transfers more suitable information for the decision-making layer. Various experiments on several benchmark datasets, including GTZAN, Ballroom, and Extended Ballroom, have verified the excellent performances of the proposed neural network. Codes and model will be available at "ttps://github.com/CaifengLiu/music-genre-classification".
1. Introduction
Music genre classification supports music information retrieval, but CNNs adapted directly from image recognition may not handle audio’s time-frequency structure or preserve useful low-level features. BBNN addresses these issues with a relatively wide, shallow architecture that carries spectrogram information from multiple scales toward the decision layer.
1. Introduction
- The paper frames genre classification as assigning a genre label to an unknown music clip, a core task in music information retrieval.Applications mentioned include music recommendation and search.Source: §1 · p. 1
- The authors contrast data-driven methods with hand-crafted features, which they say are difficult to design automatically for a specific task.Spectrograms represent audio time-frequency information and can support feature extraction from texture.Source: §1 · pp. 1–2
- The paper notes that most CNN-based music classifiers feed spectrograms into image-oriented architectures without modification.It argues that such architectures do not account for differences between spectrograms and images.Source: §1 · p. 2
- Audio genres can respond differently to time scales and frequency levels, while existing deep-learning approaches predict from the same scale.The paper argues that this motivates CNN structures that handle multiple scales of audio features.Source: §1 · p. 2
- Existing music-classification networks can lose low-level features that carry information important to genre decisions.The paper identifies preserving these features while abstracting high-level information as a central design challenge.Source: §1 · p. 2
- BBNN uses a relatively wide, shallow structure with dense connections between Inception blocks to carry low-level spectrogram information toward the decision layer.Inception blocks process multiple scales, and dense connectivity transmits their feature maps to higher layers.Source: §1 · p. 2
2. Proposed Design And Approach
BBNN’s Broadcast Module combines multi-scale Inception blocks with dense connections so feature maps from earlier blocks can reach later ones. The full network extracts local features, downsamples the expanded maps, and uses global average pooling for genre decisions.
2.1. Broadcast Module
- Inception blocks combine different convolution kernel sizes to learn features at multiple receptive fields, while dense paths transmit extracted maps to subsequent blocks.The authors describe this design as preserving feature maps for use by decision layers.Source: §2.1 · p. 3
- The Broadcast Module stacks Inception blocks with dense connections, allowing each block to receive outputs from all preceding blocks.For music genre recognition, the number of Inception blocks is fixed at three.Source: §2.1 · p. 3
- The Inception-block input concatenates the shallow-layer output with feature maps from earlier blocks, and each block applies a composite Inception function.Convolutions use mainly 1 × 1, 3 × 3, and 5 × 5 filters; the implementation adds batch normalization before each convolution.Source: §2.1 · p. 3
2.2. Network Structure
- BBNN is organized into shallow feature-extraction, Broadcast Module, transition, and decision layers.The network maps its input spectrogram to a distribution over genre labels.Source: §2.2 · pp. 3–4
- Shallow layers use a 3 × 3 convolution and 4 × 1 max pooling to extract local frequency information over a short time span.The extracted local information is passed into each Broadcast Module layer to build contextual time-frequency signatures.Source: §2.2 · p. 3
- Transition layers apply batch normalization, ReLU, a 1 × 1 convolution, and 2 × 2 average pooling to reduce feature-map size and channel count.They follow the Broadcast Module, where concatenation increases map size and channels.Source: §2.2 · p. 3
- The decision stage uses global average pooling over each feature map instead of stacked fully connected layers.The paper says this makes feature-map-to-genre correspondence easier to interpret and is less prone to overfitting.Source: §2.2 · p. 3
- Table 1 presents BBNN’s configuration, with each listed convolution layer corresponding to batch normalization, ReLU, then convolution.The stated input mel-spectrogram is 647 × 128 for 30 seconds of music.Source: Table 1 · p. 4, §2.2 · p. 4
3. Experiments
Experiments evaluate BBNN on GTZAN, Ballroom, and Extended Ballroom using mel-spectrogram inputs and 10-fold cross-validation. BBNN reports strong classification results across the datasets, including 97.1% accuracy on Ballroom and 97.2% on Extended Ballroom.
3.1. Datasets
- The three datasets include GTZAN, Ballroom, and Extended Ballroom; the extended dataset has six times more, better-quality tracks than Ballroom and is imbalanced.Ballroom has clear, constant rhythmic patterns; the Extended Ballroom imbalance poses classification challenges.Source: §3.1 · p. 4, Table 2 · p. 4
3.2. Preprocessing
- BBNN takes 647 × 128 mel-spectrograms as input; training minimizes categorical cross-entropy with ADAM and uses early stopping.Across datasets, training uses batch size 8 for 100 epochs, with the learning rate halved after three epochs without loss improvement.Source: §3.3 · p. 4, §3.2 · p. 4, Figure 3 · p. 5
3.3. Training and other details
- Evaluation averages classification accuracy across 10-fold cross-validation, with data randomly partitioned into training, testing, and validation sets in an 8/1/1 proportion.Training and validation losses converge to low values on all three datasets.Source: §3.3 · p. 4
3.4. Classification results on GTZAN
- BBNN performs slightly better than Multi-DNN on GTZAN, while the compared methods report accuracies from 85.4% to 91.0%.Most genres are correctly classified; Jazz precision and recall each reach 99%.Source: §3.4 · p. 5, Table 3 · p. 5, Table 4 · p. 5
3.5. Classification results on Ballroom
- 97.1% accuracy on Ballroom; BBNN outperforms all compared models, and its average recall rate is 0.9% higher than SOTA.Rumba and Slow Waltz are relatively easily confused, with less clear-cut genre boundaries than Cha Cha or Quickstep.Source: §3.5 · pp. 5–6, Table 5 · p. 5, Table 6 · p. 6
3.6. Classification results on Extended Ballroom
- 97.2% accuracy on Extended Ballroom; BBNN surpasses the compared CNN-based and hand-crafted methods, without pre-training on a larger dataset.Most genre recognition precisions exceed 90%, with Quickstep and Tango at 99%; Wcswing is difficult to recognize with only 0.5% of samples.Source: §3.6 · p. 6, Table 7 · p. 6
4. Conclusions and future work
BBNN is evaluated against hand-crafted and deep-learning methods on GTZAN, Ballroom, and Extended Ballroom, including the 698-track Ballroom dataset. The authors propose future work on acoustic-feature fusion and a new genre-similarity distance metric.
4. Conclusions and future work
- BBNN was evaluated on GTZAN, Ballroom, and Extended Ballroom, including the 698-track Ballroom dataset, and achieved satisfactory accuracies.The comparisons included hand-crafted feature approaches and deep-learning models with different architectures.Source: §4 · p. 6
- Future work will explore inputting acoustic features such as SSD and RH into the network through fusion strategies.Source: §4 · p. 6
- The authors also plan to use a new distance metric to compute similarity between genres at the decision-making stage.Source: §4 · pp. 6–7