Optimal delayed decisions in encoding and decoding of audio signals and general sources.

摘要

This dissertation is concerned with algorithms that optimally exploit delay for encoding or decoding decisions in certain common scenarios involving signal compression.;In applications that involve off-line encoding, such as movie-streaming over the internet, music playback from hand-held devices, and so on, the end-user is not sensitive to encoding delay. Despite this fact, encoders typically compress frame after frame of the signal, thereby restricting encoding delay. As one focus of this dissertation, delayed-decision approaches are explored, to optimize the encoding operation over the entire signal. Standards based audio-compression is chosen as the candidate setting to demonstrate the benefits of the concept. A two-layered trellis effectively optimizes both intra- and inter-frame encoding decisions while minimizing a psychoacoustically relevant distortion measure under a prescribed bit-rate constraint. The bit-stream produced is standard compatible and there is no additional decoding delay. As an accompaniment to this rate-distortion optimization paradigm, and motivated by it, modifications are proposed to the audio distortion metric itself that enhance its psychoacoustic relevance, and endeavor to enable subjectively optimal decisions.;Subsequently the focus shifts to delay at the decoder end of the compression chain. Unlike at the encoder, there are no parameter choices to make. But can the decoder, by suitable application of delay, exploit correlations if any with future frames to improve the reconstruction of the current frame? This question is particularly relevant in predictive coding scenarios, where a correlated source model is explicitly assumed. The encoder predicts the current sample from the past, and codes the prediction residual. Correlations with future samples can be exploited at the decoder end, for instance by applying a non-causal filter to smooth the regular zero-delay reconstructions. In contrast, this dissertation proposes an estimation-theoretic framework where conditional probability densities, given both past and available future information (for a fixed delay), are recursively calculated, and optimal reconstruction computed via conditional expectation. This optimal delayed decoder in turn motivates a near-optimal low complexity approximation, that employs a time-invariant look-up table or codebook approach. Applications include video compression employing motion compensated prediction, and so called 'low-delay' applications, where predictive coding is used in lieu of transform coding to avoid large framing delays and encoding complexity.