Robust predictive vector quantizer design for speech and video coding

A central component of an efficient digital audio or video communication system is the source coder. Vector quantizers (VQ) have long been used to implement efficient source coders, and gained popularity not only due to their theoretical optimality, but also due to their practical simplicity and wide applicability. Predictive VQ (PVQ) can provide significant performance gains over VQ, for audio and video sources, at the cost of a small increase in complexity. The gains are achieved by exploiting information, extracted from past samples of the source, to predict its future.; There exist three longstanding fundamental shortcomings to standard PVQ design. The first complication is due to the PVQ structure, as it utilizes a feedback loop for the prediction. Feedback may have a detrimental impact on the convergence and stability of the design procedure, as quantized residuals are fed back into the system during design, causing error accumulation. The second shortcoming is due to the piecewise constant nature of the quantizer function, which makes it difficult to optimize the predictor with respect to the overall reconstruction error. Traditionally, the predictor was simply designed to minimize the prediction error, which is potentially mismatched with the overall error. Finally, a shortcoming inherited from standard VQ design is the tendency of the design algorithm to terminate at a locally, rather than the globally, optimal solution. In this thesis, we propose an algorithm that effectively resolves all three design shortcomings simultaneously.; To eliminate the stability problems due to the feedback loop, we propose an asymptotically closed-loop design. At each iteration, we restrict the quantizer and predictor design to employ a fixed set of reconstruction vectors from the previous iteration. Clearly, the design is effectively open-loop. However, as the iterations converge, the design becomes effectively closed-loop. The approach thus has the stability of open-loop algorithms, yet asymptotically optimizes the objective closed-loop performance.; The problem due to the piecewise constant nature of the quantizer is overcome by temporarily randomizing the quantizer operation. Instead of deterministically quantizing an input vector to the nearest reproduction vector in the codebook, it is quantized to all codevectors in probability, depending on the corresponding Euclidean distance. Hence, the piecewise constant quantizer function is replaced with an effectively differentiable one, allowing joint optimization of predictor and quantizer so as to directly minimize the overall reconstruction error.; The above probabilistic framework is further exploited to resolve the third shortcoming, namely, the suboptimality of VQ design. By an appropriate gradual modification of the probability distributions, we effectively generate an annealing process capable of avoiding poor local optima. More specifically, we embed the asymptotic closed-loop technique within a deterministic annealing procedure. Substantial gains over traditional methods have been achieved in all simulations.; A further outcome of the thesis research is in the area of transmission of speech and images over packet networks. Due to the high rate of packet loss experienced in current packet networks and timely delivery constraints on streaming media, non-predictive coding techniques are commonly used. We present a new source-channel coding approach to design multi-stage VQ structures within an index interleaving framework. The approach minimizes receiver reconstruction error subject to lossy packet network communications. The robustness of the proposed system is demonstrated under various sources and packet loss conditions.