Word boundary agreementto combine multi-microphone hypotheses in distant speech recognition

In this paper we propose a technique for combining hypotheses generated in a multi-microphone setting, which exploits complementarity and collective agreement among ASR outputs of different channels. The technique draws upon the information encoded in the available set of word lattices. As a first step, we identify word boundaries in which a comprehensive inter-channel agreement is found; then, these boundaries are used to reduce the global hypothesis search space. Global word posterior probabilities are estimated for the candidate words associated to each of the bounded segments. As a result, a single combined confusion network is generated from the multiple lattices. This approach offers a novel perspective to state of the art solutions based on confusion network combination. Promising results were obtained from an experimental evaluation in a simulated domestic environment equipped with a distributed microphone network. The development and test sets were simulated using real impulse responses estimated for a large set of microphone-speaker position pairs.