Lattice-based GOP in automatic pronunciation evaluation
Yin Song, Weiqian Liang, Runsheng Liu · 2010
Goodness of pronunciation (GOP) is a confidence measure widely used in automatic pronunciation evaluation (APE) and it does not use the phone loop recognition lattice to calculate the confidence score. This paper presents a lattice-based GOP algorithm utilizing the information from the lattice, in order to obtain a more accurate confidence score. Experimental results reveal that the proposed LGOP measure generally performs better than GOP except for APE of short sentences. Moreover, the LGOP score is empirically less likely to be influenced by the length of the assessed sentences. But the calculation of the LGOP score is very time-consuming, which may be overcome by adopting more economic lattice-generating algorithms.