A Cross-Domain Approach to Temporal Envelope Shaping in Parametric Stereo Coding Using Deep Learning

In parametric stereo audio coding, at the encoder a stereo signal is downmixed to a mono signal along with a set of time-frequency dependent stereo parameters. At the decoder, using a decorrelator, a decorrelated signal is first generated from the downmix signal. A replica of the stereo signal is subsequently reconstructed based on the time-frequency dependent stereo parameters, the downmix and the decorrelated signal. A disadvantage of traditional decorrelators is that they have trouble following the temporal envelope of the mono signal due to frequency-dependent delays introduced in their processing. This is especially problematic for signals with strong, short energy bursts like transients, and leads to unwanted smearing of the decorrelated signal. In this work, we introduce a cross-domain deep learning approach for reshaping a decorrelated signal’s temporal envelope in the subband domain, making use of envelope features learned from the time-domain downmix.