RD Optimal Temporal Noise Shaping for Transform Audio Coding

In this article we investigate rate-distortion optimal temporal noise shaping for transform audio coding. Temporal noise shaping, or TNS, is a technique for reshaping the quantization noise in the time domain through open-loop linear predictive coding of frequency domain coefficients. Traditionally, a selection mechanism based on prediction gain is employed to determine whether it is advantageous to apply TNS or not. Although this method is effective for reducing coding artifacts in transient and speech signals, critical adjustment of the prediction gain threshold is necessary to avoid excessive bit rate demands. We propose the use of TNS in a rate-distortion optimization framework. Within this framework a jointly optimal selection of the prediction filter order and the quantizer for coding the coefficients can be made, such that the perceptual distortion is minimized for a given target rate. Experimental results for an MDCT-based audio coding system are presented and it is shown that TNS within an RD optimization framework outperforms the existing TNS method