A speech fluency grading system for general representation of fluency was developed to evaluate both smooth reading and more advanced skills related to content-based coherent expression and rhythmic proficiency.In addition to traditional prosodic features,this approach also analyzed pause,rhythm,links and assimilation skills in fluent English pronunciation.The scoring stage uses a hierarchical scoring fusion strategy with sentence and passage level using fusion measures such as multiple linear regression(LR),back propagation neural networks(BP),and support vector regression(SVR).The correlation and mean square errors between the ground-truth estimated scores and automatically estimated scores suggests that the non linear score fitting methods(BP,SVR) are superior to the linear method.Thus,this approach should be adopted in computer aided automatic scoring systems to improve efficiency and accuracy.