We present a method for source separation of speech and music signals when the number of sources is larger than the number of observation sensors, a problem known as underdetermined source separation. The method uses an iterative expectation-maximization procedure to estimate demixing parameters including frequencydependent attenuation and delay. To deal with noise distortion, the method treats noise explicitly as one of its parameters but identifies sources implicitly using a posteriori probabilities. We also extend the method to incorporate prior source statistics, represented as Gaussian mixture model. We evaluated the method in a set of noise conditions, and observed significant and consistent performance improvements than alternative methods.