AN INTEGRATEDSCHEME FOR ROBUST DISTRIBUTEDSPEECHRECOGNITIONOVER LOSSYPACKETNETWORKS

Inthis workwepresent acomplete setoftechniques devoted toofferrobustness against frame losses indistributed speech recognition overpacket-switched networks. Theproposed scheme iscomposed oftree techniques, twoofthemareapplied atthesender andthelast oneintherecognizer itself. Ononehand, amedia-specific Forward Error Correction (FEC)technique isusedtoallow therecovery of information within thebursts. Ontheother hand, arecognizer-based technique well knownbyits remarkable ability toreduce theeffects oflongconsecutive frame losses during recognition, theweighted Viterbi algorithm (WVA), isusedtohandle theadditional information introduced byFECcodes. Moreover, adouble stream strategy whereby interleaving canbeapplied along withFECcodes withoutanydelay increase, isalso applied. Theapplication ofinterleaving allows toreduce theperceived burst length atthereceiver, further improving therecognition performance. Asaresult, theproposed scheme canprovide anacceptable performance evenunder extremely adverse channel conditions. IndexTerms-Speech recognition, Packet switching, Channel coding, forward error correction, interleaved coding.