Adaptive noise cancelling and spectral subtraction are well-known methods for removing noise from a noise-corrupted speech. The scheme proposed in this paper combines both techniques in order to obtain large gains in the SNR at a reasonable low computational cost, The architecture of the canceller is based on a two-microphone scheme (noisy speech microphone or primary and environment noise microphone or reference). There are two main blocks to be considered: an adaptive lattice-ladder filter (ALLF) with an associated control automaton and a speech segmentation unit, and a two-channel spectral subtractor working with the residual signals from the ALLF.