Comparing Human and Machine Ensemble Width Estimation in Binaural Music Recordings under Simulated Anechoic Conditions

In recent years, there has been an increasing interest in binaural technology due to its ability to create immersive spatial audio experiences, particularly in streaming services and virtual reality applications. While audio localization studies typically focus on individual sound sources, ensemble width (EW) is crucial for scene-based analysis, as wider ensembles enhance immersion. We define intended EW as the angular span between the outermost sound sources in an ensemble, controlled during binaural synthesis. This study presents a comparison between human perception of EW and its automatic estimation under simulated anechoic conditions. Fifty-nine participants, including untrained listeners and experts, took part in listening tests, assessing 20 binaural anechoic excerpts synthesized using 2 publicly available music recordings, 2 different HRTFs, and 5 distinct EWs (0° to 90°). The excerpts were played twice in random order via headphones through a web-based survey. Only a subset of ten listeners, of which nine were experts, passed post-screening tests, with a mean absolute error (MAE) of 74.62° (±38.12°), compared to MAE of 5.92° (±0.14°) achieved a by pre-trained machine learning method using auditory modeling and gradient-boosted decision trees. This shows that while intended EW can be algorithmically extracted from synthesized recordings, it significantly differs from human perception. Participants reported insufficient externalization, front-back confusion (suggesting HRTF mismatch). The untrained listeners demonstrated response inconsistencies and a low degree of discriminability, which led to the rejection of most untrained listeners during post-screening. The findings may contribute to the development of perceptually aligned EW estimation models.