VoIP(Voice over Internet Protocol) is a technology for the voice transmission over packet networks. In real time packetized voice applications, such as the internet telephony, the packet loss is a major factor of voice quality degradation. Thus the packet loss recovery is needed to guarantee the QoS(Quality of Service).
G.729 8kbps CS-ACELP(Conjugate Structure Algebraic Code Excited Linear Prediction) coder is a representative voice codec of VoIP. A packet has voice parameters which consist of LSP(Line Spectrum Pairs), ACB(Adaptive Codebook), and FCB(Fixed Codebook) parameter. In this paper, we observe existing researches about packet loss recovery algorithms, and propose a novel algorithm to improve the performance of G.729.
Packet loss recovery algorithms are divided into two kinds, sender-based algorithms and receiver-based algorithms. Generally, receiver-based algorithms are more efficient and have a low bit transmission rate. Receiver-based algorithms are called PLC(Packet Loss Concealment). Among PLC, the algorithm by linear interpolation of parameters has a good performance. But there are some limitations. LSP and FCB parameters quantized by predictive coding are not used in the interpolation algorithm. And the voice recovered from packet loss is unnatural because parameter characteristics are not reflected in the algorithm.
To solve these problems, we propose a novel PLC using parameter interpolation with presumption and selective cubic spline interpolation. Parameter interpolation with presumption is an algorithm that presumes LSP and FCB with weighted moving average and interpolates. Selective cubic spline interpolation is an algorithm that selectively interpolates parameters by cubic spline interpolation according to the chracteristic of parameter variation.
For the verification of the proposed algorithm, comparative experiments have been performed. We measured SD(Spectral Distortion) and SNR(Signal-to-Noise Ratio) in condition of 5~30% packet loss rate, 1~3 frames/packet about 300 sentences. It is confirmed that the proposed algorithm has low SD and high SNR against the linear interpolation PLC. We also performed the preference test. We let 20 people select a better one among the existing PLC algorithm and proposed algorithm by listening test for 10 sentences. As a result, the proposed algorithm showed a higher preference score. Through the experiments, it was confirmed that the proposed PLC using parameter interpolation with presumption and selective cubic spline interpolation makes a higher quality voice recovered from packet loss effectively.
