Improving robustness of a likelihood-based beamformer in a real environment for automatic speech recognition