N-best parallel maximum likelihood beamformers for robust speech recognition

This work aims at improving speech recognition in noisy environments using a microphone array. The proposed approach is based on a preliminary generation of N-best hypotheses. The use of an adaptive maximum likelihood beamformer (the Limabeam algorithm), applied in parallel to each hypothesis, leads to an updated set of transcriptions, among which the maximally likely to clean speech models is selected. Results show that this method improves recognition accuracy over both Delay and Sum Beamforming and Unsupervised Limabeam especially at low SNRs. Results also show that it can recover the recognition errors made in the first recognition step.