A soft computing filtering approach is proposed for adaptive noise cancellation. The goal of noise cancellation is to extract the desired signal from its noise-corrupted version, using the proposed neuro-fuzzy system (NFS) as an adaptive filter. Traditional linear filtering may not be good enough to handle with the noise complexity. In the study, the NFS filter is trained in hybrid way using the well-known random optimization (RO) method and the least squares estimate (LSE) method for the noise canceling problem. The premises and the consequents of the NFS are updated for their parameters using the RO and the LSE, respectively. With the hybrid learning algorithm, the proposed approach has moderate computation and the training of the NFS filter is fast convergence. An example of noise cancellation by the proposed adaptive NFS filter is illustrated and the result is discussed. The NFS filter has stable filtering performance for noise cancellation.