Conditional Vector And Interpolation-Based Quantization For Packet Loss Resilience

Low rate compression and packet loss in real-time VOIP (Voice over IP) transimission is inevitable due to the limited channel resource especially under wireless channels. Coding with side information (CSI) is an effective way to reduce data and recover the lost signal. At the receiving end, it is the key point to describe the high quality speech by utilizing CSI efficiently with fewer bits.In this thesis, different conditional vector quantization (CVQ) codebooks are designed to realize the accuracy consistent LSF quantization, the quality control is also fulfilled by the related codebook associated with the CSI of intra-frame and inter-frame Linear Spectrum Frequency (LSF). A new CVQ scheme with linear interpolation (CVQ-I) is suggested in G.729. Furthermore, an interleaved CVQ-II for packet lost concealment is also given. Our main work is focus on:1. Estimation of the lost LSFs scheme---CVQ-I method. The intra-frame LSFs are first splitted into two parts, the low frequency component (LFC) and the high one (HFC), each are quantized by its VQ codebook trained by LBG clustering. A related codebook is created based on conditional distributions of concerning the LFC and HFC by minimizing the errors of LSF estimation. A CSI prediction with intra-frame HFC and inter-frame LFC are providing as side information to estimate the lost LFC. Experiments prove that CVQ-I with 3 redundancy bits (CVQ-I3) is better for the compromise of the storage space of codebook, bandwidth and recovery. At random loss rate 10%, the SD of our CVQ-I3 is 0.0076dB below than that of the linear interpolation (LI).2. Accuracy consistent codebook design. Different accuracy LFC, HFC and the correlated CVQ codebooks are designed by CVQ-I method. Four rates of the three codebooks are designed to meet different estimated accuracy and bandwidth.3. A lost concealment for G.729---CVQ-II. The intra-frame HFC are interleaved to the consecutive frame, thus the lost LFC can be estimated by CVQ-I. The Experiment proves that at random loss rate 10% CVQ-II is 0.03 higher in PESQ than interpolation concealment (CI).