On a new stochastic version of the EM algorithm

We present an algorithm in which the Maximisation step of the EM algorithm is replaced by a Sampling step. We describe an application of the algorithm to noise reduction for an audio signal. The results of various simulations on synthetic data are presented and compared to the results obtained using the EM algorithm and the Gibbs Sampler. A major limitation of the EM algorithm is that it can converge on local stationary points. The results we present show how our algorithm successfully overcomes this limitation.