A High-Quality General-Purpose Decorrelator with Transient Fidelity

In a multi-channel spatial audio presentation, a decorrelator is a signal processing algorithm which helps to create a diffuse sensation for the listener by minimizing any localization cues. In this thesis, the relevant psycho-acoustic and signal processing principles are reviewed, and a new decorrelator algorithm is proposed that operates blindly on a single-channel input signal and creates a 5-channel decorrelated presentation. The algorithm operates by separating transient and steady-state information in the input. The steady-state information is then split into band-pass signals approximating the critical bands of human hearing. Each output channel applies a different amount of delay to each band-pass signal with respect to the other channels. The transient information is panned randomly between the output channels with a rapidly changing location.

A listening test was conducted to determine how the new algorithm performs in comparison to two existing algorithms (drawn from the literature). The evaluation considered four attributes: clarity, naturalness, envelopment, and stability of spatial image. The test procedure was adapted from the well known MUSHRA test, and follows recommendations from research on evaluating up-mix algorithms. The large variance in many variables of the data made it difficult to achieve statistical significance; however, some specific conclusions were made on the relative performance of the algorithms.

Block Diagram of Algorithm:

Layout of Random 5 Channel Panning:
Panning Layout

Audio Examples (48 kHz, stereo wav):


(citations for source material are in complete thesis document, p. 41-42)