Correcting State Error by Making Full Use of Late Packet for Prediction-Based Decoding in VoIP

Prediction-based speech codecs are known to be sensitive to packet loss because of inter-frame dependencies in their prediction states. If a packet is lost, internal states between the encoder and decoder will become automatically out of synchronization. This causes the state error propagation and will degrade the speech quality even if the decoder receives the following packets correctly. Within the context of voice transmission over packet networks, some packets arrive too late to be decoded and have to be given up. A new method is proposed to use late packets to correct the state error of the decoder, which minimize the state error propagation. The state tracking mechanism of the method can ensure the decoder's making full use of late packets without increasing the overall end-to-end delay. Simulation results show that the speech quality is greatly improved with this method at the price of additional complexity.