Temporal Noise Shaping on MDCT Subband Signals for Transform Audio Coding

A common problem when coding transient signals in transform audio codecs is the temporal unmasking of quantization noise leading to artifacts such as pre-echoes. In this work, a new method for temporal shaping of quantization noise is introduced. Subband merging/splitting with overlapping windows is applied on the MDCT (Modified Discrete Cosine Transform) spectrum in the encoder and decoder. This allows direct modification of the temporal subband signals for different frequency ranges in a way that results in a flattening of the temporal envelopes in the encoder before quantization and a reshaping to their original envelope at the decoder. The reshaping process at the decoder not only restores the envelopes of the original subband signals, but also shapes the quantization noise accordingly, which results in quantization noise temporally shaped similarly to the original subband signal. This avoids temporal unmasking of quantization noise and the pre-echo artifact. For the temporal flattening/reshaping process the application of gains transmitted via Autoregressive model as well as a low side information approach via companding are proposed and evaluated regarding their noise shaping capabilities in an experimental codec system.