Music Retrieval by Rhythmic Similarity Applied on Greek and African Traditional Music

Iasonas Antonopoulos, Aggelos Pikrakis, Sergios Theodoridis, Olmo Cornelis, Dirk Moelants, Marc LemanISMIR 2007 · OpenReview

Abstract

This paper presents a robust methodology for retrieving traditional Greek and African music based on rhythmic similarity. Moving beyond Western-centric models that rely on standard 4/4 meters and stable tempi, the proposed system extracts "rhythmic signatures" through self-similarity analysis of chroma-based features. By comparing these signatures using Dynamic Time Warping (DTW), the researchers successfully classified complex, improvisational music with high accuracy, demonstrating the system's effectiveness for cultural heritage preservation and ethnomusicology.

Introduction

Most Music Information Retrieval (MIR) research focuses on Western popular music, which typically follows standard meters and stable tempi. Traditional music from Greece and Africa presents a significant challenge: it features complex meters (e.g., 7/8), intricate interval systems, and a high degree of improvisation. Existing methods often struggle to capture these nuances. This research aims to bridge the gap by creating a system that identifies similar tracks based on their rhythmic "feel" or signature, providing a vital tool for musicologists.

Methodology

The authors propose a multi-step pipeline to transform raw audio into a comparable rhythmic fingerprint:

  1. Feature Extraction: The system utilizes "chroma-based MFCCs." Unlike standard Mel-Frequency Cepstral Coefficients, these are tuned to a chromatic scale (36 filters over six octaves), making them more sensitive to the tonal structures inherent in traditional music.
  2. Dimensionality Reduction: Singular Value Decomposition (SVD) is applied to the feature sequence to retain only the most significant components, effectively reducing noise.
  3. Rhythmic Signature Extraction: This is the core innovation. The audio is segmented into 4-second windows, and a Self-Similarity Matrix (SSM) is calculated for each. By averaging values along the diagonals of these matrices, the system produces a rhythmic signature ($R_\mu$). Valleys in this signature represent periodicities—the "heartbeat" of the meter and tempo.
Comparison of rhythmic signatures for a 7/8 meter song (blue) versus a 2/4 meter song (red). The distinct patterns allow for effective discrimination between different rhythmic structures.
  1. Similarity Measurement: To compare signatures, the system uses Dynamic Time Warping (DTW) with Sakoe-Chiba constraints. DTW is ideal for this application as it can "stretch" or "compress" signals to find the best match, accounting for slight variations in performance speed.

Data and Evaluation

The study utilized two distinct corpora of non-Western music:

  • Greek Traditional Dance Corpus: 220 tracks categorized into four genres based on meter and tempo (e.g., 2/4, 3/4, 7/8).
  • African Traditional Music Corpus: 103 field recordings from the Royal Museum of Central-Africa, split into test sets based on meter (3/4 to 6/4) and prototypical rhythmic patterns (quintuple, sextuple, etc.).

Key Results

The method demonstrated impressive classification performance across both datasets. For the Greek music corpus, the system achieved precision and recall rates mostly above 90% across all four classes. It was particularly adept at distinguishing the complex 7/8 meter from standard Western meters.

Precision and Recall tables for the Greek Traditional Dance Corpus, showing high accuracy across all four rhythmic classes.

In the African dataset, precision for meter classification ranged from 64% to 92.6%, with the 6/4 meter identified most accurately. When classifying by specific rhythmic patterns, the system achieved precision between 60% and 86.7%. Errors typically occurred when performers added or omitted percussive events, creating variants that the system occasionally confused.

Discussion and Conclusion

A key insight of this work is that self-similarity serves as an excellent proxy for meter in improvisational contexts. Instead of attempting difficult beat detection, the "valleys" in the rhythmic signature naturally emerge at intervals corresponding to the musical meter. Furthermore, by aligning the filter bank with the chromatic scale, the features better represent tonal structures, making rhythmic repetitions clearer.

While the results are promising, the authors noted limitations in automated "thumbnailing" (finding representative segments) for dense percussive structures. Future work will explore extracting multiple rhythmic signatures per track to account for pieces that change rhythm or tempo mid-way, and scaling the system to larger, more diverse music corpora.

Summary updated

BibTeX
@inproceedings{antonopoulos2007music,
  author = {Iasonas Antonopoulos and Aggelos Pikrakis and Sergios Theodoridis and Olmo Cornelis and Dirk Moelants and Marc Leman},
  title = {{Music Retrieval by Rhythmic Similarity Applied on Greek and African Traditional Music}},
  booktitle = {ISMIR},
  year = {2007},
  url = {https://openreview.net/forum?id=xyAauwMHeag},
}
Figure