Multi-microphone signal processing techniques have the potential to greatly improve the robustness of speech recognition (ASR) in distant microphone settings. However, in everyday environments, typified by complex non-stationary noise backgrounds, designing effective multi-microphone speech recognition systems is non trivial. In particular, optimal performance requires the tight integration of the front-end signal processing and the back-end statistical speech and noise source modelling. The best way to achieve this in a modern deep learning speech recognition framework remains unclear. Further, variability in microphone array design --- and consequent lack of real training data for any particular configuration --- may mean that systems have to be able to generalise from audio captured using mismatched microphone geometries or produced using simulation.