Packet Loss Concealment for Predictive Speech Coding Based on Extrapolation of Speech Waveform

Most packet loss concealment (PLC) algorithms for predictive speech coders are based on extrapolating the excitation signal. This paper presents an alternative algorithm that extrapolates the speech waveform directly. The algorithm avoids the additional delay associated with the overlap-add operation in the ITU-T G.711 Appendix I by utilizing the ";ringing"; of the synthesis filter. It also performs special operations to update the filter states to prepare the predictive decoder for the next good frame. This algorithm achieves robust PLC performance when applied to the BroadVoicereg standard coders. It can also be applied to other predictive speech coders.