An error-protected speech recognition system for wirelesscommunications

摘要

Future wireless multimedia terminals will have a variety ofnapplications that require speech recognition capabilities. We consider anrobust distributed speech recognition system where representativenparameters of the speech signal are extracted at the wireless terminalnand transmitted to a centralized automatic speech recognition (ASR)nserver. We propose two unequal error protection schemes for the ASR bitnstream and demonstrate the satisfactory performance of these schemes forntypical wireless cellular channels. In addition, a "soft-feature" errornconcealment strategy is introduced at the ASR server that usesn"soft-outputs" from the channel decoder to compute the marginalndistribution of only the reliable features during likelihood computationnat the speech recognizer. This soft-feature error concealment techniquenreduces the ASR error rate by more than a factor of 2.5 for certainnchannels. Also considered is a channel decoding technique with sourceninformation that improves ASR performance