Robust recognition using multiple utterances

Increasing the reliability of the results of automatic speech recognition systems is an important research and development issue. Although recent systems have been shown to achieve quite high recognition results for limited tasks, this may not be good enough for some applications, where some bits of information are critical, and have to be recognized correctly. In this paper, we suggest a method for the improvement of the robustness of speech recognition, using a speaker independent HMM, of either a word, phrase, or a full sentence, by taking advantage of repeated utterances of the same content. The method can be applied to most con gurations of HMM based recognizers, and does not require additional model training. Recognition experiments showed improvement in recognition accuracy, under quiet and noisy conditions.