Yes, this is noisy. You might be able to bring down the noise by using a larger pool of normal samples to build the reference profile (instead of T/N pairing), and maybe exclude problematic genomic regions if you know what they are.