Due to the finite bandwidth of practical wireless systems, one multipath component can manifest itself as a discrete pulse consisting of multiple taps in the digital delay domain. This effect is called channel leakage, which complicates the multipath delay estimation problem. In this paper, we develop a new algorithm to estimate multipath delays of leaked channels by leveraging the knowledge of pulse-shaping functions, which can be used to support fine-grained WiFi sensing applications. Specifically, we express the channel impulse response (CIR) as a linear combination of overcomplete basis vectors corresponding to different delays. Considering the limited number of paths in physical environments, we formulate the multipath delay estimation as a sparse recovery problem. We then propose a sparse Bayesian learning (SBL) method to estimate the sparse vector and determine the number of physical paths and their associated delay parameters from the positions of the nonzero entries in the sparse vector. Simulation results show that our algorithm can accurately determine the number of paths, and achieve superior accuracy in path delay estimation and channel reconstruction compared to two benchmarking schemes.