Perceptual spectrum quantization

We propose two new quantization schemes, based on a novel speech spectrum distortion measure with interframe memory, to illustrate and exploit the perceptual importance of the time evolution of spectrum vectors. The proposed methods punish quick changes in the time evolution of the spectrum error. This leads to a slowly evolving quantized spectrum, with a maintained ability to track quick transitions in the unquantized spectrum. We found both the methods to give an increased subjective performance, compared to standard systems operating at the same rate. Listening tests point at possible gains of around 2 bits per frame.