# Benchmarking Sub-Genre Classification For Mainstage Dance Music
## Abstract
Music classification, a cornerstone of music information retrieval, supports a wide array of applications. To address the lack of comprehensive datasets and effective methods for sub-genre classification in mainstage dance music, we introduce a novel benchmark featuring a new dataset and baseline. Our dataset expands the scope of sub-genres to reflect the diversity of recent mainstage live sets performed by leading DJs at global music festivals, capturing the vibrant and rapidly evolving electronic dance music (EDM) scene that engages millions of fans worldwide. We employ a continuous soft labeling approach to accommodate tracks blending multiple sub-genres, preserving their inherent complexity. Experiments demonstrate that even state-of-the-art multimodal large language models (MLLMs) struggle with this task, while our specialized baseline models achieve high accuracy. This benchmark supports applications such as music recommendation, DJ set curation, and interactive multimedia systems, with video demos provided. Our code and data are all open-sourced at https://github.com/Gariscat/housex-v2.git.
## 1 Introduction
Being one of the most fundamental and important tasks in music information retrieval (MIR), music genre classification has mainly focused on broad genres [1, 2, 3, 4]. Despite continuous progress, existing datasets lack fine-grained labels that can capture the nuances within electronic dance music (EDM), and with 0/1 labels, they hardly manifest the category overlap in EDM data. Moreover, current universal models have subpar performance in specific tasks [5, 6, 7, 8]. The limitations are evident in the contexts of mainstage DJ sets, where tracks usually fall into sub-genres within the broad category of house music. Hence, the unique challenges of EDM necessitate specialized datasets and algorithms that are tailored to its structural characteristics and the complexity of its production techniques.
To address this gap, we introduce a new benchmark specifically targeting at the classification of house music sub-genres. A dataset is designed to provide annotations from a list of 8 sub-genres. These 8 classes were selected to represent a comprehensive range of prominent sub-genres, reflecting their popularity and distinctive characteristics that highlight the diversity and evolution of EDM shaped by the influence of music festivals and digital media. Unlike existing dataset HouseX [9], we introduce soft labeling instead of 0/1 categorical labeling to provide detailed and nuanced representation of the music. In addition, a baseline model is presented using spectral features [10, 11], building a foundation for future research in dance music genre classification. Lastly, we develop a prototype of automated music visualization based on our proposed model, demonstrating its potential to enhance the visual experience.
Overall, this work aims to advance MIR for mainstage dance music by offering data, models, and a demo application. Our key contributions include: (a) improved annotation paradigm with an extended dataset; (b) strong specialized models and experiment results where MLLMs fail; (c) a multimedia application demo (visual FX automation).
Table 1: Determinants of sub-genres, where $x$ - $y$ indicates the lowerbound $x$ and the upperbound $y$ of the attribute, voted by human musicians/producers.
| Genre | Lead Inst. | Chord Inst. | Bass Inst. | Groove | Rhythm | Distortion | Organicity |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Progressive House | ✓ | ✓ | ✓ | 4-5 | 1-3 | 1-2 | 3-5 |
| Future House | ✓ | $∼$ | ✓ | 3-4 | 4-5 | 3-4 | 1-3 |
| Bass House | ✓ | ✗ | ✓ | 2-4 | 3-4 | 4-5 | 1-2 |
| Tech House | ✗ | ✗ | ✓ | 3-5 | 3-4 | 2-4 | 1-3 |
| Deep House | ✓ | $∼$ | ✓ | 2-3 | 1-2 | 1-2 | 2-4 |
| Bigroom | ✓ | ✗ | ✗ | 1-2 | 1-3 | 4-5 | 1-2 |
| Future Rave | ✓ | ✗ | ✓ | 4-5 | 3-5 | 2-4 | 1-2 |
| Slap House | ✓ | ✗ | ✓ | 3-4 | 3-5 | 1-3 | 1-2 |
- ✓ denotes “yes”, ✗ denotes “no”, and $∼$ denotes “uncertain”, which could be either yes or no depending on the track.
<details>
<summary>images/pipeline-pdfv8.jpg Details</summary>
### Visual Description
## Diagram: Audio Dataset Production Workflow and Baseline Architecture
### Overview
The diagram illustrates two interconnected components:
1. **Dataset Production Workflow** (left section)
2. **Baseline Architecture** (right section)
Both sections emphasize feature extraction, transformation, and machine learning processes for audio data.
---
### Components/Axes
#### (a) Dataset Production Workflow
1. **Input**:
- Raw Audio (labeled as "1 Raw Audio")
- Labels (labeled as "3 Label")
2. **Processing Steps**:
- **Manual Annotation**: Human-labeled audio segments.
- **Drop Segmentation**: Rule-based segmentation by volume thresholds (e.g., "Drop 1", "Drop i", "Drop n").
- **Time-to-Frequency Transformations**:
- `mel-spectrogram`
- `cqt-chromagram`
- `vqt-chromagram`
- **Feature Stacking**: Combines transformed features into "Clip Feature".
3. **Output**:
- Labeled dataset with segmented audio clips and extracted features.
#### (b) Baseline Architecture
1. **Input**:
- Clip Features (from dataset production workflow).
2. **Feature Extraction Module**:
- **CNN or ViT**: Processes sub-clip features.
- **Transformer Encoder**:
- **Multi-Head Attention**: Captures contextual relationships.
- **Feed Forward**: Non-linear transformations.
- **Trainable Modules**: Learnable parameters for feature refinement.
3. **Output**:
- **Label Prediction**: Final classification output (`y_i`).
- **MLP Last-Layer Feature Visualization**: Visualizes intermediate features as clusters (e.g., green, blue, red dots).
---
### Detailed Analysis
#### Dataset Production Workflow
- **Drop Segmentation**: Segments raw audio into 7.5-second clips (e.g., "Clips 1", "Clips j").
- **Feature Extraction**:
- `mel-spectrogram`: Captures frequency magnitude over time.
- `cqt-chromagram`: Represents pitch-class content.
- `vqt-chromagram`: Visualizes harmonic structure.
- **Stacking**: Combines multiple features (e.g., mel, CQT, VQT) into a unified "Clip Feature".
#### Baseline Architecture
- **Transformer Encoder**:
- **Multi-Head Attention**: Processes sub-clip features in parallel, capturing long-range dependencies.
- **Feed Forward**: Applies non-linear transformations to attention outputs.
- **MLP**: Maps encoded features to label predictions (`y_i`).
- **Feature Visualization**:
- **Sub-Clip Features**: Represented as 2D grids (e.g., 128x128).
- **Clip Features**: Stacked 3D tensors with dimensions:
- `hop length*` (time resolution)
- `clip length` (frequency resolution)
- `number of mel-banks` (spectral bins).
---
### Key Observations
1. **Integration of Techniques**:
- Combines traditional audio features (mel-spectrogram, CQT) with deep learning (transformer, MLP).
2. **Hyper-Parameters**:
- `hop length*`: Controls time resolution in feature extraction.
- `clip length`: Defines frequency resolution.
3. **Rule-Based Segmentation**:
- Segments audio based on volume thresholds, ensuring consistent clip lengths.
4. **Visualization**:
- MLP last-layer features are clustered (e.g., green, blue, red), suggesting class separation.
---
### Interpretation
1. **Dataset Production**:
- The workflow emphasizes preprocessing raw audio into structured, labeled clips with multi-modal features. This ensures robustness for downstream tasks.
2. **Baseline Architecture**:
- The transformer encoder and MLP highlight the importance of contextual modeling and non-linear transformations for accurate label prediction.
3. **Feature Visualization**:
- The clustered visualization of MLP features suggests effective feature discrimination, critical for model interpretability.
4. **Hyper-Parameter Tuning**:
- Adjusting `hop length` and `clip length` balances temporal and spectral resolution, impacting model performance.
This diagram underscores a pipeline that bridges raw audio data with advanced machine learning, leveraging both traditional signal processing and deep learning techniques.
</details>
Figure 1: Pipeline Structure from Data Collection to Model Design
## 2 Related Work
Music information retrieval (MIR) has been a research focus since the early 21st century, as evidenced by several comprehensive surveys [12, 13, 14]. Traditional music genre datasets, like GTZAN [1], FMA [2] and MSD [15], often focus on broad genres like pop, country, and rock. Other general datasets like MusicCaps [16] provide music-text pairs from which one can extract genre information. The HouseX dataset [9] advanced classification in EDM but faced challenges like limited category richness and scale. Prior works [17, 18] studied EDM sub-genre classification by traditional methods and deep learning, respectively. In addition, previous works tackling drop In electronic dance music, the “drop” refers to a structural climax, typically marked by an abrupt increase in sonic intensity, rhythmic complexity, and spectral energy, serving a role comparable to the chorus in popular music. detection [19, 20, 21] also highlight the emerging research interest in MIR for EDM. Alongside traditional deep learning methods, recent progress in language models for music understanding like [22] and [23], and strong multimodal large language models (MLLMs) like [5, 24, 7, 8] could also be used to classify music audio.
This work presents a new dataset expanding from a smaller dataset [9], featuring 1000+ production-level tracks from top record labels that span across 8 sub-genres. To the best of our knowledge, given that [18] only open-sourced their model checkpoints, our work is the only publicly available large-scale dataset for EDM sub-genre classification. We use soft labeling to capture tracks with multiple sub-genre traits Please refer to the code repository for an example.. Our architecture employs convolutional neural networks (CNNs) [25, 26, 27] and vision transformers (ViTs) [28] to extract features with a sliding window, followed by some transformer encoder layers to predict the target distribution. Such customized models for sub-genre classification significantly outperform the existing MLLMs that take audio as direct input.
## 3 Methodology
This section outlines the methodology used in our study, encompassing the definition and annotation of the dataset, as well as the training of the classification model. The complete pipeline, from data preparation to model training/inference, is illustrated in Figure 1. For simplicity and clarity, we refer to sub-genres of house music as “genres” throughout this section unless otherwise specified.
### 3.1 Dataset
EDM encompasses a diverse range of sub-genres, each characterized by distinct rhythmic, timbral, and structural features. While these sub-genres are well-known within the EDM community, they may be unfamiliar to broader audiences in audio signal processing and music information retrieval. To facilitate understanding of our dataset and classification methodology, we first define the eight EDM sub-genres considered in this study, highlighting their key characteristics and relevance to our analysis, followed by the collection and labeling process.
#### 3.1.1 Definitions
We provide both qualitative and quantitative descriptions of the 8 genres. A key observation is that the values of soft labels are not absolute per se. Even for professionals, one track can be interpreted in multiple ways. For example, track $X$ is generally classified as genre $A$ but also exhibits characteristics of genre $B$ . Therefore, a soft label can better preserve this nuance than a binary 0/1 label that assigns the track solely to genre $A$ . In Table 1, we show how different sub-genres are characterized by the following aspects.
1. The first three columns (xxx. Inst.) indicate whether the track specifically assigns instruments for the parts, respectively. For example, some future house tracks do not use chord instruments since their lead sounds can cover the mid frequencies.
1. Groove measures the energy fluctuation of the instrumental part (no percussion) within a beat, from 1 (very present, extremely short attack) to 5 (highly-sidechained).
1. Rhythm discusses how fragmented/off-beat the notes are. We consider the length of notes played by the instruments in the track. 1 means extremely coarse (mainly 1/4 notes and 1/8 notes) and 5 means extremely fine (1/32 notes exist).
1. Distortion describes the extent to which the track uses “dirty” and “inharmonic” sounds/effects, such as clipping, flanging, etc. 1 is very clean (hardly any noticeable distortion) and 5 is very harsh (almost distortion in every bar).
1. Organicity measures the extent to which the track uses organic instruments like acoustic piano/guitar/strings, etc.
We also provide plain-English descriptions, including genre-specific characteristics (e.g., vocal leads, plucked bass) and one representative example for each sub-genre. Due to space constraints, these are included in the code repository.
#### 3.1.2 Annotation
The input data are represented as triplets, each comprising the raw audio ( $X_raw$ ), its extracted acoustic features ( $X_feat$ ), and a corresponding label ( $Y$ ), as illustrated by numbers from 1 to 3 in Figure 1. Note that our model does not take in audio waveforms, which are used as input by MLLMs such as Qwen-Audio [5] and Kimi-Audio [8] that include built-in spectrogram extractors or other audio encoders like Whisper [29].
Our dataset comprises over 1,000 selected tracks sourced from renowned international record companies. Dataset statistics are presented in Table 2. These tracks are stored as uncompressed raw audio ( $X_raw$ ). For each track, we only consider the drop, which is the most representative segment of the style. At times, listening to other sections like the intro could provide a rough estimate of the category, but it is the drop that determines the exact genre of a song. To identify the drop, we detect excerpts where the volume consistently remains above a threshold $V_margin$ is a hyper-parameter set as 1.5dB empirically. ( $V_thres=V_max-V_margin$ ), where $V_max$ denotes the maximum volume within the audio. Note that before detection, some rule-based smoothing process is applied to mitigate loudness fluctuations. After identifying the segments that meet the aforementioned criteria, several 7.5-second clips (around 4 bars We extended the clip length which was 1 bar in HouseX, since usually 4 bars make up a loop that better exhibits the style.) are randomly sampled from each segment, ensuring a comprehensive representation of the track.
Subsequently, feature extraction is performed on each clip. Using librosa [30], we compute: mel-spectrograms ( $X_mel$ ), CQT-chromagrams ( $X_cqt$ ), VQT-chromagrams ( $X_vqt$ ), which are then stacked to produce the final audio feature matrix ( $X_feat$ ). Detailed hyper-parameters for generating the mel-spectrogram, CQT, and VQT, are comprehensively shown in our code repository.
All tracks in the dataset are manually annotated with a total of eight distinct labels. The first four labels align with those previously defined by the HouseX dataset [9]. To enhance classification granularity to a level suitable for representing most contemporary mainstage live sets played in music festivals, we introduce four additional genres. Recognizing that certain tracks may exhibit characteristics of multiple genres, we employ soft labeling techniques to add more information to our labels, which gives better performance than hard 0/1 labeling, as shown in Sec. 4.
Table 2: An overview of the dataset with the size for each category along with the sample rate and the BPM (beats per minute) span.
| Genre | # of Clips | |
| --- | --- | --- |
| Train. | Val. | |
| Progressive House | 1215 | 344 |
| Future House | 1192 | 348 |
| Bass House | 1102 | 120 |
| Tech House | 643 | 200 |
| Deep House | 591 | 116 |
| Bigroom | 774 | 284 |
| Future Rave | 920 | 108 |
| Slap House | 667 | 232 |
| Total | 7104 | 1752 |
| # of Tracks (Train. & Val.) | 1035 | |
| Sample Rate | 44100 Hz | |
| BPM Range | 115-130 | |
Table 3: Experiment results on the validation set, where the precision, the recall and the F1 score are all weighted across the 8 categories.
| Architecture | Precision $⋆$ | Recall $⋆$ | F1 $⋆$ | | | | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| chroma info $†$ | w | w/o | w | w/o | w | w/o | | | | | | |
| label setup $‡$ | hard | soft | hard | soft | hard | soft | hard | soft | hard | soft | hard | soft |
| Ours (ViT_B_16) | $0.077$ | $0.039$ | $0.039$ | $0.039$ | $0.224$ | $0.199$ | $0.199$ | $0.199$ | $0.109$ | $0.066$ | $0.066$ | $0.066$ |
| Ours (VGG11_BN) | $0.746$ | $0.751$ | $0.739$ | $0.765$ | $0.716$ | $0.748$ | $0.735$ | $0.757$ | $0.722$ | $0.747$ | $0.731$ | $0.756$ |
| Ours (DenseNet201) | $0.730$ | $0.767$ | $0.786$ | 0.790 | $0.712$ | $0.749$ | $0.752$ | $0.743$ | $0.716$ | $0.741$ | $0.757$ | $0.751$ |
| Ours (ResNet152) | $0.760$ | $0.761$ | $0.776$ | $0.780$ | $0.720$ | $0.737$ | $0.748$ | 0.761 | $0.718$ | $0.744$ | $0.756$ | 0.764 |
| Qwen-Audio-Chat § | 0.122 | 0.099 | 0.037 | | | | | | | | | |
| Qwen2-Audio-7B-Instruct § | 0.202 | 0.129 | 0.031 | | | | | | | | | |
| Qwen2.5-Omni-7B § | 0.008 | 0.075 | 0.014 | | | | | | | | | |
| Kimi-Audio-7B-Instruct § | 0.121 | 0.187 | 0.109 | | | | | | | | | |
| MU-LLaMA § | 0.050 | 0.224 | 0.081 | | | | | | | | | |
| MusiLingo § | 0.257 | 0.283 | 0.116 | | | | | | | | | |
- The 3 metrics here are weighted averages across all 8 genres.
- This row indicates whether we include chromagrams in the data representation.
- This row indicates whether a hard 0/1 labeling or a soft labeling approach is used in the data representation.
- We use GPT-4o-generated doc and human-producer expertise in the text prompt for these models, provided in the code repository.
<details>
<summary>x1.png Details</summary>
### Visual Description
## 3D Scatter Plots: Dimensionality Reduction Visualization
### Overview
The image contains three 3D scatter plots comparing the performance of three dimensionality reduction techniques (PCA, t-SNE, UMAP) on a dataset of music genre classifications. Each plot visualizes data points distributed across three principal components, with color-coded clusters representing distinct music genres.
### Components/Axes
1. **Plots**:
- **Left**: PCA (Principal Component Analysis)
- **Center**: t-SNE (t-Distributed Stochastic Neighbor Embedding)
- **Right**: UMAP (Uniform Manifold Approximation and Projection)
2. **Axes**:
- All plots share the same axis labels:
- **X-axis**: Component 1
- **Y-axis**: Component 2
- **Z-axis**: Component 3
- Axis ranges vary slightly:
- PCA/t-SNE: -20 to +20
- UMAP: -5 to +15 (X), -5 to +15 (Y), -2.5 to +12.5 (Z)
3. **Legend**:
- Located at the bottom center of the image
- Color-coded classes (10 genres):
- Progressive House (green)
- Future House (purple)
- Bass House (orange)
- Techno (yellow)
- Deep House (blue)
- Bigroom (pink)
- Future Rave (brown)
- Slap House (gray)
- No explicit numerical values or uncertainty ranges provided in the legend.
### Detailed Analysis
1. **PCA Plot**:
- Data points are widely dispersed across all three components.
- Progressive House (green) and Future House (purple) form the largest clusters, occupying the upper-left quadrant.
- Slap House (gray) and Bigroom (pink) appear as smaller, isolated clusters in the lower-right quadrant.
2. **t-SNE Plot**:
- Points show moderate clustering with some overlap between genres.
- Techno (yellow) and Deep House (blue) form a dense cluster near the center.
- Future Rave (brown) and Bass House (orange) are positioned in the lower-left quadrant.
3. **UMAP Plot**:
- Tightest clustering observed, with distinct separation between genres.
- Progressive House (green) dominates the upper-right quadrant.
- Slap House (gray) forms a compact cluster in the lower-left corner.
- UMAP's Z-axis (Component 3) shows the most pronounced separation between classes.
### Key Observations
1. **Cluster Density**:
- UMAP demonstrates the highest cluster separation (e.g., Slap House in lower-left, Progressive House in upper-right).
- PCA shows the least distinct clustering, with significant overlap between genres.
2. **Axis Utilization**:
- In PCA, Component 1 and 2 dominate the spread, while Component 3 has minimal variation.
- t-SNE utilizes all three components more evenly.
- UMAP's Z-axis (Component 3) shows the greatest discriminative power.
3. **Color Consistency**:
- All plots maintain identical color coding per legend (e.g., pink = Bigroom in all plots).
- No discrepancies observed between legend labels and data point colors.
### Interpretation
The visualization demonstrates how different dimensionality reduction techniques preserve data structure:
1. **PCA** (linear method) struggles to capture non-linear relationships, resulting in dispersed clusters.
2. **t-SNE** balances global and local structure preservation, showing moderate clustering with some overlap.
3. **UMAP** excels at maintaining both global topology and local relationships, producing the most distinct genre separation.
Notable patterns:
- Progressive House consistently appears in the upper-left quadrant across all methods, suggesting it has dominant features in the first two principal components.
- Slap House (gray) shows the most consistent positioning in the lower-left quadrant, indicating shared characteristics across reduction methods.
- The Z-axis (Component 3) becomes increasingly important for discrimination in t-SNE and UMAP, highlighting its role in capturing subtle genre distinctions.
This analysis suggests UMAP may be the most effective technique for this dataset, while PCA's linear limitations become apparent in the visualization.
</details>
Figure 2: 3D Feature Visualization of the Training Set after Dimension Reduction
### 3.2 Model
Given a dataset of audio features $\{X_feat^i\}_i=1^N$ with corresponding soft labels $p(y_i|X_feat^i)$ that represent probability distributions over possible genres, our objective is to train a neural network, parameterized by $θ$ , to predict genre distributions $q_θ(y_i|X_feat^i)$ that closely match these soft labels.
To achieve this goal, we minimize the Kullback-Leibler (KL) divergence between the true distribution $p(y_i|X_feat^i)$ and the predicted distribution $q_θ(y_i|X_feat^i)$ :
$$
KL(p\parallel q_θ)=∑_i=1^N∑_y_{i∈Y}p(y_i|X_feat^i)\log≤ft(\frac{p(y_i|X_feat^i)}{q_θ(y_i|X_feat^i)}\right). \tag{1}
$$
Discarding the part that does not depend on $θ$ , this is equivalent to minimizing the cross-entropy loss:
$$
L(θ)=-∑_i=1^N∑_y_{i∈Y}p(y_i|X_feat^i)\log≤ft(q_θ(y_i|X_feat^i)\right). \tag{2}
$$
Note that this formulation is consistent with cross-entropy for hard labels where $p$ simply becomes the indicator function that indicates the sole genre $X_feat^i$ belongs to.
The architecture of network $θ$ is inspired by AST (audio spectrogram transformer) [31] trained on the AudioSet [32]. As shown in column (b) in Figure 1, we derive a sequence of overlapping patches $M$ is determined by the hyper-parameter “hop length”. $\{X_feat^i,j\}_j=1^M$ from each acoustic feature matrix $X_feat^i$ , where $X_feat^i,j$ is square with shape ( $3× 224× 224$ ). Then each square feature patch is fed into a vision encoder, which essentially consists of conventional architectures used for image classification [25, 26, 27, 28]. The output for the sequence of input feature patches is a corresponding sequence of embeddings $\{Z_emb^i,j\}_j=1^M$ . Positional encodings are then added to this sequence of embeddings, which are subsequently passed through several stacked transformer [33] encoder layers. The first (on the $seq\_len$ axis) output from this transformer is finally fed into a linear layer to obtain the predicted distribution.
Compared with the previous architecture in HouseX [9], which is solely an image classifier, we convert the raw feature matrix into a sequence of patches to accommodate the quadrupled clip length. In fact, keep resizing the feature matrix into squares in a brute-force manner would lead to severe information loss on the temporal axis, which is crucial for characterizing genres like future house that contain very fragmentary music notes or samples.
## 4 Results
This section presents the experiment results alongside our interpretations and findings. Table 3 provides numerical details and Figure 2 shows the final embedding space after feature reduction by PCA, t-SNE [34] and UMAP [35]. By convention, precision, recall, and F1-score were computed after sharpening the soft labels into hard categories (e.g., converting a soft label of [ $0.25,0.75,0,…$ ] into [ $0,1,0,…$ ]) for evaluation purposes. This approach allows the results to be interpreted as the model’s ability to correctly identify the primary characteristic of a song.
Four popular CNN/ViT architectures (with comparable number of parameters) serve as the feature extractor. Experiments show that all our CNN-based specialized models outperform the rest significantly. For now, the lower performance of ViT-based methods is attributed to the need of high temporal resolution for EDM genre classification, where CNN architectures inherently provide stronger inductive biases suited for audio spectrogram data. Again, our purpose is to justify the necessity of building specific dataset rather than “beat” LLM-based models. In other words, MLLMs are quite likely to perform well when provided with proper post-training data; otherwise, they struggle to align textual background information with domain-specific audio features. In addition, models trained on soft labels perform uniformly better than those trained from (sharpened) 0/1 labels, which supports our claim that soft labeling provides richer information of the tracks. Moreover, models trained on composite data with chromagrams fail to surpass those trained solely on mel-spectrogram. We attribute this phenomenon to the domain gap between the RGB space of ImageNet [36] and the mel-CQT-VQT space. Future work may explore refining network architectures or leveraging existing music encoders–such as MERT [37], MusicFM [38], or MuQ [39] –which are trained on broader musical genres, to improve performance on our composite dataset.
We also extract the embeddings before the final linear layer and provide visualization with dimension reduction techniques. The figures indicate that progressive house, bigroom and slap house are relatively well-separated, aligning with our annotations where most non-0/1 labels involve other genres.
For hardware configuration, at least 24G GPU memory is required to complete one run with a batch size of 4. We assign one NVIDIA RTX A6000 GPU to each data setting (whether or not to use chromagram and whether or not to use soft labels).
## 5 Applications
We propose some real-world scenarios for such classification algorithm. A straightforward application is music recommendation system tailored for listeners with specific preferences on certain sub-genres. Besides, this algorithm can boost productivity in multimedia contexts, such as automated MV generation and visuals generation given certain pre-defined rules [9]. For illustrative purposes, we prototyped a visual automation system simulated in Blender 3D Demos are available in the links given in the code repository..
## 6 Conclusion
We developed a comprehensive house music classification benchmark with 8 classes to address the limitations of existing datasets, such as limited sub-genre representation and inter-class overlaps. Our dataset employs continuous soft labeling to better capture track characteristics. Our study highlights the inadequacy of existing machine learning methods in understanding sub-genres of electronic music, thus calling for the development of specialized datasets and models.
Future work may explore several promising directions. One avenue is to further scale up the dataset to better utilize the CQT and VQT feature spaces. Since labeling large datasets is impractical for the machine learning community alone, a collaborative approach with producers/songwriters from the music industry is recommended. Additionally, future MIR research should also extend beyond sub-genre classification to encompass timbral and rhythmic characteristics, which retrieves more information that could benefit downstream scenes. On the other hand, training an MLLM capable of captioning EDM tracks with descriptive attributes remains a challenging but promising task to empower downstream applications.
## REFERENCES
- [1] G. Tzanetakis and P. Cook, “Musical Genre Classification of Audio Signals,” IEEE Transactions on Speech and Audio Processing, vol. 10, no. 5, pp. 293–302, 2002.
- [2] M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “FMA: A Dataset for Music Analysis,” in 18th International Society for Music Information Retrieval Conference (ISMIR), 2017. [Online]. Available: https://arxiv.org/abs/1612.01840
- [3] M. Defferrard, S. P. Mohanty, S. F. Carroll, and M. Salathé, “Learning to Recognize Musical Genre from Audio,” in The 2018 Web Conference Companion. ACM Press, 2018. [Online]. Available: https://arxiv.org/abs/1803.05337
- [4] M. Zhou, S. Xu, Z. Liu, Z. Wang, F. Yu, W. Li, and B. Han, “CCMusic: An Open and Diverse Database for Chinese Music Information Retrieval Research,” Transactions of the International Society for Music Information Retrieval, vol. 8, no. 1, p. 22–38, 2025. [Online]. Available: http://dx.doi.org/10.5334/tismir.194
- [5] Y. Chu, J. Xu, X. Zhou, Q. Yang, S. Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models,” arXiv preprint arXiv:2311.07919, 2023.
- [6] Y. Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y. Leng, Y. Lv, J. He, J. Lin, C. Zhou, and J. Zhou, “Qwen2-Audio Technical Report,” 2024. [Online]. Available: https://arxiv.org/abs/2407.10759
- [7] J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, K. Chen, J. Wang, Y. Fan, K. Dang, B. Zhang, X. Wang, Y. Chu, and J. Lin, “Qwen2.5-Omni Technical Report,” 2025. [Online]. Available: https://arxiv.org/abs/2503.20215
- [8] KimiTeam, “Kimi-Audio Technical Report,” 2025. [Online]. Available: https://arxiv.org/abs/2504.18425
- [9] X. Li, “HouseX: A Fine-Grained House Music Dataset and Its Potential in the Music Industry,” in 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2022, pp. 335–341.
- [10] C. Schörkhuber and A. Klapuri, “Constant-Q transform toolbox for music processing,” in 7th Sound and Music Computing Conference (SMC), 2010, pp. 3–64.
- [11] C. Schörkhuber, A. Klapuri, N. Holighaus, and M. Doerfler, “A Matlab Toolbox for Efficient Perfect Reconstruction Time-Frequency Transforms with Log-Frequency Resolution,” in Proceedings of the AES International Conference, 01 2014.
- [12] R. Typke, F. Wiering, and R. C. Veltkamp, “A Survey of Music Information Retrieval Systems,” in Proceedings of the International Conference on Music Information Retrieval, 01 2005, pp. 153–160.
- [13] M. Schedl, E. Gómez, and J. Urbano, Music Information Retrieval: Recent Developments and Applications. Now Foundations and Trends, 2014.
- [14] F. Simonetta, S. Ntalampiras, and F. Avanzini, “Multimodal Music Information Processing and Retrieval: Survey and Future Challenges,” in 2019 International Workshop on Multilayer Music Representation and Processing (MMRP). IEEE, Jan. 2019. [Online]. Available: http://dx.doi.org/10.1109/MMRP.2019.00012
- [15] T. Bertin-Mahieux, D. P. Ellis, B. Whitman, and P. Lamere, “The Million Song Dataset,” in Proceedings of the 12th International Society for Music Information Retrieval Conference, 2011, pp. 591–596.
- [16] A. Agostinelli, T. I. Denk, Z. Borsos, J. Engel, M. Verzetti, A. Caillon, Q. Huang, A. Jansen, A. Roberts, M. Tagliasacchi, M. Sharifi, N. Zeghidour, and C. Frank, “MusicLM: Generating Music From Text,” 2023. [Online]. Available: https://arxiv.org/abs/2301.11325
- [17] A. Caparrini, J. Arroyo, L. Pérez-Molina, and J. Sánchez-Hernández, “Automatic subgenre classification in an electronic dance music taxonomy,” Journal of New Music Research, vol. 49, no. 3, pp. 269–284, 2020. [Online]. Available: https://doi.org/10.1080/09298215.2020.1761399
- [18] W.-H. Hsu, B.-Y. Chen, and Y.-H. Yang, “Deep Learning Based EDM Subgenre Classification using Mel-Spectrogram and Tempogram Features,” 2021. [Online]. Available: https://arxiv.org/abs/2110.08862
- [19] G. Argüello, L. A. Lanzendörfer, and R. Wattenhofer, “Cue Point Estimation using Object Detection,” 2024. [Online]. Available: https://arxiv.org/abs/2407.06823
- [20] K. Yadati, M. A. Larson, C. C. S. Liem, and A. Hanjalic, “Detecting Drops in Electronic Dance Music: Content based approaches to a socially significant music event,” in Proceedings of the 15th International Society for Music Information Retrieval Conference, ISMIR 2014, Taipei, Taiwan, October 27-31, 2014, H. Wang et al., Eds., 2014, pp. 143–148. [Online]. Available: http://www.terasoft.com.tw/conf/ismir2014/proceedings/T026_297_Paper.pdf
- [21] Y.-S. Huang, S.-Y. Chou, and Y.-H. Yang, “Pop Music Highlighter: Marking the Emotion Keypoints,” Transactions of the International Society for Music Information Retrieval, vol. 1, no. 1, p. 68–78, 2018. [Online]. Available: http://dx.doi.org/10.5334/tismir.14
- [22] S. Liu, A. S. Hussain, C. Sun, and Y. Shan, “Music Understanding LLaMA: Advancing Text-to-Music Generation with Question Answering and Captioning,” 2023. [Online]. Available: https://arxiv.org/abs/2308.11276
- [23] Z. Deng, Y. Ma, Y. Liu, R. Guo, G. Zhang, W. Chen, W. Huang, and E. Benetos, “MusiLingo: Bridging Music and Text with Pre-trained Language Models for Music Captioning and Query Response,” in Proceedings of the 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2024). Association for Computational Linguistics, 2024.
- [24] Y. Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y. Leng, Y. Lv, J. He, J. Lin, C. Zhou, and J. Zhou, “Qwen2-Audio Technical Report,” 2024. [Online]. Available: https://arxiv.org/abs/2407.10759
- [25] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2015. [Online]. Available: https://api.semanticscholar.org/CorpusID:206594692
- [26] G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely Connected Convolutional Networks,” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2261–2269, 2016. [Online]. Available: https://api.semanticscholar.org/CorpusID:9433631
- [27] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” CoRR, vol. abs/1409.1556, 2014. [Online]. Available: https://api.semanticscholar.org/CorpusID:14124313
- [28] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” 2021. [Online]. Available: https://arxiv.org/abs/2010.11929
- [29] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust Speech Recognition via Large-Scale Weak Supervision,” 2022. [Online]. Available: https://arxiv.org/abs/2212.04356
- [30] Brian McFee, Colin Raffel, Dawen Liang, Daniel P.W. Ellis, Matt McVicar, Eric Battenberg, and Oriol Nieto, “librosa: Audio and Music Signal Analysis in Python,” in Proceedings of the 14th Python in Science Conference, Kathryn Huff and James Bergstra, Eds., 2015, pp. 18 – 24.
- [31] Y. Gong, Y.-A. Chung, and J. Glass, “AST: Audio Spectrogram Transformer,” in Proc. Interspeech 2021, 2021, pp. 571–575.
- [32] J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio Set: An ontology and human-labeled dataset for audio events,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 776–780.
- [33] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need,” in Proceedings of the 31st International Conference on Neural Information Processing Systems, ser. NIPS’17. Red Hook, NY, USA: Curran Associates Inc., 2017, p. 6000–6010.
- [34] L. van der Maaten and G. Hinton, “Visualizing Data using t-SNE,” Journal of Machine Learning Research, vol. 9, no. 86, pp. 2579–2605, 2008. [Online]. Available: http://jmlr.org/papers/v9/vandermaaten08a.html
- [35] L. McInnes, J. Healy, and J. Melville, “UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction,” 2020. [Online]. Available: https://arxiv.org/abs/1802.03426
- [36] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255.
- [37] Y. Li, R. Yuan, G. Zhang, Y. Ma, X. Chen, H. Yin, C. Lin, A. Ragni, E. Benetos, N. Gyenge, R. Dannenberg, R. Liu, W. Chen, G. Xia, Y. Shi, W. Huang, Y. Guo, and J. Fu, “MERT: Acoustic Music Understanding Model with Large-Scale Self-supervised Training,” 2023.
- [38] M. Won, Y.-N. Hung, and D. Le, “A Foundation Model for Music Informatics,” 2023. [Online]. Available: https://arxiv.org/abs/2311.03318
- [39] H. Zhu, Y. Zhou, H. Chen, J. Yu, Z. Ma, R. Gu, Y. Luo, W. Tan, and X. Chen, “MuQ: Self-Supervised Music Representation Learning with Mel Residual Vector Quantization,” 2025. [Online]. Available: https://arxiv.org/abs/2501.01108