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

Performance of distant-talking speech recognizers in real noisy environments can be increased using a microphone array. In this work we propose an N-best extension of the Limabeam algorithm, which is a likelihood-based adaptive lter-and-sum beamformer. We show that this algorithm can be used to optimize the noisy acoustic features using in parallel the N-best hypothesized transcriptions generated at a rst recognition step. The parallel and independent optimizations increase the likelihood of minimal word error rate hypotheses and the resulting N-best hypotheses list is automatically re-ranked. Results show improvements over delay-and-sum beamforming and Unsupervised Limabeam on a real database with considerable amount of noise and limited reverberation.