AES Europe 2026

Audio event-classification models trained on AudioSet are
widely adopted; form a central component of the state of
the art in machine listening, yet their behavior when
deployed in complex, open acoustic environments remains
largely unexplored. In this study, we evaluate several
widely adopted AudioSet-pretrained
architectures—particularly models from the PANNs family,
including MobileNetV2; Wavegram; Transformer-based
PaSST model—when applied to a real operational scenario at
the commercial Port of Valencia, Spain. We observed a
recurring; systematic unexpected behavior: the models
frequently assigned disproportionately high probability to
the class Music for non-musical industrial;
transportation sounds. These mislabeled events included
train-wheel squealing, motorcycle acceleration, emergency
sirens,; reversing beeps—sound categories that are
common in port logistics environments but acoustically
different from music. By analyzing the probability
distributions output by the models, we demonstrate that
this erroneous Music activation is not an isolated failure
but a pervasive pattern across several architectures. Our
findings highlight a critical gap in the robustness;
domain generalization of AudioSet-derived models;
emphasize the need for targeted adaptation techniques when
deploying them in real industrial settings.