Phone-dependent transformation of posterior probability measure for automatic pronunciation quality evaluation

Abstract

Posterior probability measure is widely accepted as the most promising feature for automatic pronunciation quality evaluation. However, this measure is not phonetically consistent. This work presents a novel trainable phone-dependent transformation of posterior probability to deal with the problem. Both linear and non-linear transforms are investigated. Close form solution is found for linear transformation and gradient-based method is derived for nonlinear transformation. Experimental results on the database of 3685 people showed significant improvement. The cross-correlation between human and machine scores increases from 0.582 to 0.760.