21st Century Electronica: MIR Techniques for Classification and Performance

Dimitri Diakopoulos, Owen Vallis, Jordan Hochenbaum, Jim W. Murphy, Ajay KapurISMIR 2009 · OpenReview

Introduction

Modern Music Information Retrieval (MIR) provides powerful tools for analyzing audio, yet a significant gap remains between these techniques and the interfaces used by live DJs. Traditional digital DJing software often mimics physical turntables or relies on text-based lists, failing to fully utilize the rich data provided by modern classification and similarity clustering. This paper introduces "DJen," a system that leverages MIR for subgenre classification and uses Self-Organizing Maps (SOMs) to create an intuitive, multi-touch tabletop interface for live electronic music performance.

Feature Extraction and Classification

The researchers developed a pipeline to classify electronica subgenres by extracting 8-dimensional feature vectors using the ChucK audio programming language. Features were categorized into three groups: spectral (centroid, flux, rolloff), loudness (RMS, panning, panning delta), and temporal (bass onsets, inter-onset time).

A key finding was the significance of panning; unlike acoustic music, electronica relies heavily on artificial stereo manipulation as a stylistic marker. The authors found that including panning and temporal features improved classification accuracy by up to 7.7% on their 6-class dataset, proving that production style is a vital genre identifier in electronic music.

Classification Performance

Using the Weka environment, several machine learning models were tested, including k-Nearest Neighbor (IBk), Decision Trees (J48), Multi-Layer Perceptrons (MLP), and Support Vector Machines (SMO). The k-NN classifier (IBk) proved most effective across the board.

Classification accuracy across different machine learning models for 6-class and 4-class datasets.

On a 6-class dataset consisting of 600 clips across genres like IDM, House, and Techno, the system achieved 75.2% accuracy. When highly overlapping genres were removed to create a 4-class set, accuracy rose to 87.0%. Interestingly, the "confusion" observed in the models—such as between IDM and Drum and Bass—often mirrors the sonic similarities inherent in these genres, which both feature complex breakbeats.

The DJen Interface

The high-dimensional audio data is reduced to a 2D space using a Self-Organizing Map (SOM), which clusters musically similar tracks together. This map is visualized on the "Bricktable," a multi-touch surface that allows for a tangible user experience.

A user interacting with the DJen application on a multi-touch tabletop, drawing a path to generate a playlist.

Instead of searching through linear lists, DJs can browse the music library spatially. By drawing a continuous path across the map with a finger, the system automatically generates a playlist that smoothly transitions through sonic styles. This gesture-based approach allows for "happy accidents" and collaborative performance, where two DJs can draw paths simultaneously to create a mathematically averaged set-list that blends their tastes in real-time.

Conclusion

The DJen system demonstrates that SOMs combined with multi-touch interfaces offer a viable, low-latency paradigm for live music selection. By shifting from text-based search to spatial interaction, the system empowers DJs to make more intuitive and creative choices. Future work aims to incorporate automatic beat-matching and expanded collaborative features, further bridging the gap between algorithmic analysis and creative live performance.

Summary updated

BibTeX
@inproceedings{diakopoulos2009st,
  author = {Dimitri Diakopoulos and Owen Vallis and Jordan Hochenbaum and Jim W. Murphy and Ajay Kapur},
  title = {{21st Century Electronica: MIR Techniques for Classification and Performance}},
  booktitle = {ISMIR},
  year = {2009},
  url = {https://openreview.net/forum?id=o_Nlf3HyxUY},
}
Figure