Hypothesis combination for Slovak dictation speech recognition

Combination of multiple speech recognition systems is the most used method for improving speech recognition accuracy. The combination is performed at the feature level or the systems are exchanging informations between each other during decoding process or are combined afterwards using their outputs in form of N-best hypothesis or lattices. This paper provides initial experiments with system combination for Slovak language speech recognition using well known combination tool, the Recognition Output Voting Error Reduction (ROVER) from National Institute of Standards and Technology (NIST). Two kinds of scores provided to ROVER are here explored. The first one is based on normalized posterior probabilities and the second one on confidence scores of words in recognized sentence. Also new method for improving the efficiency of combination by smoothing the scores is presented.