<p>Non-contact sound signal analysis has gained increasing attention in mechanical fault diagnosis. Recent studies indicate that methods inspired by human auditory perception show promise in improving diagnostic performance. This paper proposes a gear fault diagnosis method based on Bark scale feature extraction, leveraging the close relationship between the Bark scale and human hearing. Specifically, the time-varying specific loudness model is employed to compute the loudness energy of each subband in the Bark domain of gear sound signals. Features are then extracted from subbands with higher loudness energy to reduce redundancy. A support vector machine classifier is used for fault classification. Experimental results show that the lowest accuracy recorded for our method using a single microphone was still 98.92%, significantly outperforming conventional feature extraction approaches. Moreover, it attains performance comparable to the multi-microphone data feature fusion method at a lower computational cost. The method also demonstrates strong generalization capability when applied to bearing fault diagnosis, achieving high diagnostic accuracy.</p>