The intrinsic complexity associated with passive sonar data makes the task of target recognition extremely challenging. The conventional classifier architectures based on hand-engineered feature transforms often fail miserably to disentangle the high-dimensional non-linear structures in the observed target records. Although the modern deep learning algorithms through hierarchical feature learning yield acceptable success rates, they often require tremendous amounts of data when trained in a supervised manner. An unsupervised generative framework utilizing a variational autoencoder (VAE) is proposed in this work in order to create better disentangled representations for the downstream classification task. The disentanglement is further enforced using a hyperparameter <inline-formula><tex-math notation="LaTeX">\beta </tex-math></inline-formula>. For the purpose of better segregating the spectro-temporal features, an intermediate non-linearly scaled time-frequency representation is also employed in conjunction with <inline-formula><tex-math notation="LaTeX">\beta </tex-math></inline-formula>-VAE. Experimental analysis of various classifier configurations yields encouraging results in terms of data efficiency and classification accuracy on target records collected from various locations of the Indian Ocean.