SVM-based lost packets concealment for ASR applications over IP

Voice over IP is becoming very popular due to the huge range of services that can benefit from integrating different media (voice, audio, data, etc.). Besides, voice-enabled interfaces for those services based on speech recognition are being very actively researched. Nevertheless, as it has been already pointed out ([8]), the impact of packet losses on speech recognizers is very significant. In this paper, we have compared the usual concealment method (typically implemented by speech codecs) with an SVM-based one. The proposed method is substantially better in terms of mean-square error when isolated packet losses are considered. On the other hand, preliminary ASR experiments in a more realistic environment (bursty packet losses) show just a slight enhancement of the recognition accuracy. We expect significant improvements by adapting the SVM-based method to the bursty losses.