Two distinct fields of research into robust speech recognition are the use of microphone arrays for signal enhancement and the use of independent frequency sub-band models for robust recognition. In this article, we propose and investigate the integration of these two techniques on two different levels. First, a broad-band beamforming microphone array allows for natural integration with sub-band speech recognition as the beamformer is implemented as a combination of band-limited sub-arrays. Rather than recombining the sub-array outputs to give a single enhanced output, we fuse the output of separate hidden Markov models trained on each sub-array frequency band. Second, a dynamic sub-band weighting algorithm is proposed in which the cross- and auto-spectral densities of the microphone inputs are used to estimate the reliability of each frequency band. The proposed multi-channel sub-band system is evaluated on an isolated digit recognition task and compared to both a standard full-band microphone array system and a single channel sub-band system.