Adaptive detection of CPFSK signals corrupted by Class-A impulse noise is presented. The EM algorithm (E for expectation, M for maximize) is used to estimate the noise characteristics. Performance of the EM algorithm and the adaptive detector are reported. The results show that the adaptive detector performs very close to optimal for reasonable sample sizes and well-separated components. Even when the sample size is reduced and the components are close together, the performance of the adaptive detector is still very acceptable despite the degradation. The EM algorithm is not restricted to class-A noise.