An Application of Empirical Mode Decomposition on Tempo Induction from Music Recordings

Aggelos Pikrakis, Sergios TheodoridisISMIR 2007 · OpenReview

Introduction

Extracting tempo—the perceived rate of the musical beat—from raw audio is a cornerstone of Music Information Retrieval (MIR). While humans naturally synchronize to music, automated tempo induction remains challenging due to the dynamic and non-stationary nature of audio signals. Standard algorithms often struggle with non-percussive transitions or complex rhythmic structures. This paper explores a novel approach using Empirical Mode Decomposition (EMD), a technique typically reserved for geophysical or biological signal processing, to isolate local rhythmic oscillations and improve tempo estimation accuracy.

Methodology

The proposed pipeline treats tempo induction as a multi-stage signal processing and clustering problem. The core innovation lies in the use of Self-Similarity Matrices (SSMs) and EMD to extract a robust "rhythmic signature."

  1. Feature Extraction: The audio is processed in 5-second overlapping windows. For each window, the energy envelope is used to construct an SSM. The mean value of each diagonal in the SSM is calculated to produce a 1D signal, $B(k)$, termed the rhythmic signature. In this signal, local minima (valleys) represent periodicities inherent in the audio.
The rhythmic signature $B(k)$ extracted from SSM diagonals, where periodicities manifest as local minima.
  1. EMD-Based Clustering: To ensure the algorithm focuses on stable rhythmic sections, EMD is applied to decompose each signature into Intrinsic Mode Functions (IMFs). Signatures are clustered based on their IMF profiles (number of modes and energy distribution). This step effectively filters out "noisy" segments like vocal-only bridges or non-rhythmic introductions.
  2. Tempo Induction: For each stable cluster, the mean signature is decomposed via EMD. The algorithm analyzes the two highest-energy IMFs within a standard tempo range (60–220 bpm). Valleys are scored based on their mathematical relationship to smaller lags (subdivisions), and the highest-scoring periodicity determines the cluster's tempo. A global track tempo is then derived from a smoothed histogram of these cluster-level estimates.

Key Results

The method was evaluated on a dataset of 400 recordings spanning Western Pop/Rock and Traditional Greek Folk music. The algorithm achieved an overall accuracy of 79% in inducing the notated tempo within a 3.5% margin of error.

Analysis of the failure cases revealed that the majority were "octave errors," where the algorithm identified a tempo exactly double or half the notated value. This is a common phenomenon in MIR, as rhythmic subdivisions (like 1/8th notes) can sometimes exhibit stronger periodic signatures than the primary beat (1/4 notes). In Greek Folk dances, which often feature complex meters, the algorithm occasionally locked onto the measure lag rather than the beat lag.

Breakdown of tempo induction errors, showing the prevalence of octave-related misestimations.

Discussion and Future Work

The study demonstrates that EMD is a natural fit for rhythmic analysis because it algorithmically operates on local extrema, aligning with how periodicities manifest in SSM diagonals. By separating fast oscillations from slow ones, EMD naturally disentangles different rhythmic levels (e.g., separating 1/8th note pulses from 1/4 note beats).

Future research will focus on relaxing the current strict clustering constraints—which currently utilize only about 25% of the audio data—to improve robustness. Additionally, the authors aim to extend the framework to handle non-binary meters (e.g., 7/8 or 5/4) and investigate the correlation between algorithmic output and human perceived tempo, which can vary from strict musical notation.

Summary updated

BibTeX
@inproceedings{pikrakis2007application,
  author = {Aggelos Pikrakis and Sergios Theodoridis},
  title = {{An Application of Empirical Mode Decomposition on Tempo Induction from Music Recordings}},
  booktitle = {ISMIR},
  year = {2007},
  url = {https://openreview.net/forum?id=AhotSP7-T51},
}
Figure