Burst-like packet loss compensation using a DTW/HMM hybrid scheme

Packet loss in IP and mobile networks leads to DSR systems performance degradation. Several robustness approaches have been proposed to compensate the effect of packet loss. Many of these techniques rely on temporal correlation between successive voice frames to recover data and have demonstrated improvements for minimal packet losses. Other techniques have been applied in decoding stage to control over the unfavorable influence of missing frames. In this paper, we propose a DTW/HMM hybrid scheme to deal with long burst-like packet loss. For each word-level HMM, DTW is used to align the partially corrupted data with a representative utterance. Then corresponding frames to the missing part are suggested for data reconstruction. Results have shown significant improvements over comparable methods for a Persian isolated-word recognition task in noisy conditions and when up to 50% of frames are lost.