Quantifying perceptual distortion in scalably compressed MPEG audio

A scalably compressed bitstream is one which can be streamed and decoded at a wide variety of bitrates, and it is therefore compatible with communications channels of varying capacity. The audio coding portions of the MPEG 2 and 4 standards support fine-grained scalability through the use of bit slice arithmetic coding (BSAC). Human subjective analysis of BSAC, however, has shown that it performs poorly at low bitrates; seemingly random tonal patterns are superimposed on the actual audio. Here, we develop a new approach for objectively characterizing such distortion and validate it with human subjective trials. Unlike most other objective performance metrics, the proposed approach does not require sample-accurate sequence synchronization. As a comparison, we also apply the ITU-R BS.1387-1 objective testing recommendation to the same audio sequences and quantify how well it predicts the observed subjective quality.