We present a joint spatial and spectral denoising front-end for Track 1 of the 2nd CHiME Speech Separation and Recognition Challenge based on the Flexible Audio Source Separation Toolbox (FASST). We represent the sources by nonnegative matrix factorization (NMF) and full-rank spatial covariances, which are known to be appropriate for the modeling of small source movements. We then learn acoustic models for automatic speech recognition (ASR) on the enhanced training data. We obtain 40% average error rate reduction due to speech separation compared to multicondition training alone.