Low complexity rate-distortion optimized time-segmentation for audio coding

In this paper, we investigate a reduced complexity approach to rate-distortion optimized time-segmentation in audio coding. Instead of the conventional closed-loop approach for determining the coding distortions, they are estimated from a set of features extracted from the audio signal. Care is taken to ensure that properties such as convex and non-increasing rate-distortion curves carry over from the training data to the estimated rate-distortion pairs. With computational complexity reductions of a factor close to 10, perceptual listening tests reveal a slight reduction of the signal quality, while maintaining a large improvement over fixed segmentation.