Abstract
This thesis discusses the details of research conducted in order to contribute towards the development of a generic model that predicts multichannel audio quality. The review in this thesis evaluated the existing objective models that predict audio quality. It was concluded from the review that most objective models that exists today are not capable of predicting multichannel audio quality in their current form. Therefore, important multichannel audio quality attributes were identified and an attempt was made to predict some of them using features derived from the recordings themselves. The project was completed in two phases. The selected attributes in the first phase were basic audio quality, timbral fidelity, frontal spatial fidelity and surround spatial fidelity since they were the most frequently reported. Envelopment was selected in the second phase since it was reported as an important attribute of multichannel audio in several elicitation experiments. The listening tests in the first phase were conducted according to ITU-R BS 1534-1 recommendation. A novel test paradigm was employed for evaluating envelopment. The models were calibrated by employing regression analysis techniques. The models in the first phase were of double-ended type and features IACC measurements, spectral centroid, spectral rolloff and centroid of coherence were proved to be useful for the predictions. The model for predicting envelopment was of single-ended type and features IACC measurements, spectral rolloff, area of sound distribution, inter-channel coherence and extent of coverage angle was proved to be important for prediction. The calibrated models were validated using the scores obtained from independent listening tests. The predicted scores from validation experiments showed high correlation with the actual scores and the accuracy of the models were comparable to the inter-listener errors encountered in typical listening tests. The developed models could either be used as independent applications or act as building blocks of a generic model that predicts multichannel audio quality.