COST278 and ISCA Tutorial and Research Workshop (ITRW) on Robustness Issues in Conversational Interaction
University of East Anglia, Norwich, UK
In this paper we propose a new method for error protection in speech recognition over IP networks. In such scenario, one of the main problems that should be addressed is the problem of packet losses. Several error concealment techniques that have been proposed, like frame repetition or interpolation, have shown to be very efficient when losses appear in short bursts, but their effectiveness is seriously degraded when longer bursts happen. In this work we propose the use of a new Viterbi weighted scheme for improving recognition in this situation. The basic idea is to use a different weighting factor for each component in the feature vector, which is time-varied using an estimation of the statistic correlation as a guide.
Bibliographic reference. Cardenal-López, Antonio / García-Mateo, Carmen (2004): "Correlation based soft-decoding for distributed speech recognition over IP networks", In Robust2004, paper 27.