Feature Extraction for the Prediction of Multichannel Spatial Audio Fidelity

This paper seeks to present an algorithm for the prediction of frontal spatial fidelity and surround spatial fidelity of multichannel audio, which are two attributes of the subjective parameter called basic audio quality. A number of features chosen to represent spectral and spatial changes were extracted from a set of recordings and used in a regression model as independent variables for the prediction of spatial fidelities. The calibration of the model was done by ridge regression using a database of scores obtained from a series of formal listening tests. The statistically significant features based on interaural cross correlation and spectral features found from an initial model were employed to build a simplified model and these selected features were validated. The results obtained from the validation experiment were highly correlated with the listening test scores and had a low standard error comparable to that encountered in typical listening tests. The applicability of the developed algorithm is limited to predicting the basic audio quality of low-pass filtered and down-mixed recordings (as obtained in listening tests based on a multistimulus test paradigm with reference and two anchors: a 3.5-kHz low-pass filtered signal and a mono signal)