This talk will present some speech recognition results obtained with
two microphone array speech corpora, the noisy CMU multi-microphone
corpus and an artificially created reverberant array database based on
the CSR-I (WSJ0) corpus. I will discuss the influence of noise and
reverberation time on word error rates obtained when testing single
channel data with a standard ASR System trained on clean or
reverberant data. A second line of investigation considers the
potential usefulness of employing several channels in the recognition
process. Various experiments demonstrate the effects of
(delay-and-sum) beamforming, as well as the degree of agreement and
complementarity between recognizer transcriptions based on different
channels.