Assessing Models for Estimation Ensemble Width in Binaural Music Recordings: Robustness to Reverberation and Noise

Authors

  • Paweł Antoniuk Białystok University of Technology, Faculty of Computer Science
  • Sławomir Krzysztof Zieliński Białystok University of Technology, Faculty of Computer Science

Abstract

Binaural technology has been known for decades. However, advancements in software and consumer electronics have facilitated its widespread adoption, primarily in the postmillennium era. As binaural sound becomes more popular, the demand for spatial analysis tools is expected to grow. This paper evaluates three methods for assessing ensemble width in binaural music recordings: (1) an auditory model with decision trees, (2) a neural network model, and (3) a spatial spectrogram approach. Under ideal, anechoic conditions, the auditory model performed best with a mean absolute error (MAE) of 6.59° (±0.11°), followed by the neural network (8.57° ±0.19°) and the technique based on spatial spectrograms (13.54° ±0.92°). Extending previous work, this study analyzes the methods’ robustness to reverberation and noise. Noise resilience tests indicate moderate resistance, with the auditory model yielding an MAE of 12.34° at a 10 dB signal-to-noise ratio. However, reverberation tests show a significant drop in accuracy even at an RT60 reverberation time of 0.1 seconds. The findings may contribute to the improvement of models for estimating ensemble width in binaural recordings of music, which could influence the development of binaural sound analysis tools, with potential applications in audio production.

Additional Files

Published

2025-10-13

Issue

Section

Acoustics