HouseX: A Fine-Grained House Music Dataset and Its Potential in the Music Industry

Xinyu Li

TL;DR

HouseX is a new, specialized dataset designed to move beyond broad "Electronic" music labels by classifying house music into four distinct sub-genres: Future, Bass, Progressive, and Melodic. By focusing on the "drop" of a track and utilizing computer vision models on mel-spectrograms, the researchers achieved high classification accuracy and demonstrated how these models can automate real-time music visualization for the industry.

What Problem Does This Paper Address?

In the world of Music Information Retrieval (MIR), most existing datasets are "coarse-grained." They can tell the difference between Jazz and Rock, but they struggle with the nuances of electronic music. For a DJ or a music recommendation engine, labeling a track simply as "Electronic" is unhelpful—it’s like labeling every dish in a restaurant as "Food."

The challenge is that many electronic sub-genres share the same tempo (BPM) and time signature (usually 4/4). To distinguish between them, a model must understand timbre and texture. This paper introduces HouseX to fill this gap, providing a dataset that captures the subtle stylistic differences that define modern house music sub-genres.

Methods

The researchers treated audio classification as a computer vision problem. The pipeline follows three main steps:

  1. Slicing: Audio tracks are cut into 1.875-second slices (equivalent to one bar at 128 BPM).
  2. Conversion: These slices are transformed into mel-spectrograms—visual representations of the spectrum of frequencies over time, adjusted to match human hearing.
  3. Classification: These "images" are fed into various neural network architectures.

The team tested several "heavyweight" models (ResNet18, VGG16, DenseNet121) against "lightweight" mobile-friendly models (MobileNetV3, ShuffleNetV2) and a Vision Transformer (ViT). Notably, they also manually annotated the "drop" (the climax) of every track, creating a specific "drops-only" subset to see if the most energetic part of a song provides a clearer stylistic fingerprint.

Datasets

The paper introduces the HouseX dataset:

  • Composition: 160 tracks (40 tracks each for Future House, Bass House, Progressive House, and Melodic House).
  • Annotations: Includes sub-genre labels and precise timestamps for the start and end of the "drop."
  • Comparison: The authors contrast HouseX with legacy datasets like GTZAN (10 general genres) and FMA (8 general genres), noting that all tracks in HouseX would simply be lumped into a single category in those older benchmarks.

Key Insights and Ideas

  • The "Drop" is the Key: Models performed significantly better when trained and tested only on the "drop" sections of tracks. This confirms the intuition that the climax of a house track is where its sub-genre characteristics (like the "dirty" bass of Bass House or the "supersaw" leads of Progressive House) are most pronounced.
  • Timbre Requires Depth: Larger, deeper models like VGG16 and ResNet18 outperformed lightweight models. This suggests that the features defining a sub-genre are "high-level" and complex, requiring the stronger representational power of deeper networks.
  • The "Future vs. Melodic" Dilemma: The models (and even human listeners) often confuse Future House and Melodic House. This is because both often use "happy" chord progressions and bouncy, synthesized leads, highlighting the inherent subjectivity and overlap in music taxonomy.
  • Automated Aesthetics: By mapping sub-genres to specific color palettes (e.g., Bass House = Black/White/Purple), the authors showed that AI can handle the "vibe" of a live set, allowing for automated lighting and visuals that react to the music's style in real-time.

Key Results

  • Top Performers: VGG16 and ResNet18 emerged as the strongest models.
  • Accuracy Boost: On the full dataset, VGG16 achieved a test accuracy of 71.48%. However, when focusing specifically on the drops-only subset, the accuracy jumped to 81.37%.
  • Sub-genre Distinction: Using the "drops-only" data, the models could distinguish between "cool" genres (Future/Bass) and "emotional" genres (Progressive/Melodic) with nearly 95% accuracy.

Limitations and Future Work

The primary limitation is the dataset size; 160 tracks is a solid start but small for deep learning. The authors also noted that while the models are good, PCA analysis shows that the feature space between certain sub-genres (like Future and Bass house) still has significant overlap.

Future work aims to expand the dataset to include other popular styles like Tropical House and to move beyond just color automation. The researchers envision a system where AI controls the actual geometry and displacement of 3D objects in a virtual environment, creating a fully immersive, automated sci-fi music visualizer.

Summary updated

Abstract

Machine sound classification has been one of the fundamental tasks of music technology. A major branch of sound classification is the classification of music genres. However, though covering most genres of music, existing music genre datasets often do not contain fine-grained labels that indicate the detailed sub-genres of music. In consideration of the consistency of genres of songs in a mixtape or in a DJ (live) set, we have collected and annotated a dataset of house music that provide 4 sub-genre labels, namely future house, bass house, progressive house and melodic house. Experiments show that our annotations well exhibit the characteristics of different categories. Also, we have built baseline models that classify the sub-genre based on the mel-spectrograms of a track, achieving strongly competitive results. Besides, we have put forward a few application scenarios of our dataset and baseline model, with a simulated sci-fi tunnel as a short demo built and rendered in a 3D modeling software, with the colors of the lights automated by the output of our model.
BibTeX
@inproceedings{li2022housex,
  author = {Xinyu Li},
  title = {{HouseX: A Fine-Grained House Music Dataset and Its Potential in the Music Industry}},
  booktitle = {2022},
  year = {2022},
  doi = {10.23919/apsipaasc55919.2022.9980316},
}

Loading related content...

Figure