Mathematical methods generate scores of a dissimilarity or distance measure that is used to evaluate the closeness of the input feature vector (a learner's speech) to a set of predefined feature vectors (normalized native speech). However, the technique of pattern comparison is of great interest for pronunciation evaluation. The Dynamic Time Warping (DTW) algorithm, which is an instance of the general class of algorithms known as dynamic programming and one of the oldest and most major ones in speech recognition, was used at the previous works[18].
However, recently, HMMs are more widely used because they provide a natural and much more reliable way of recognizing speech for a wide range of applications and integrate well into CALL systems geared toward training L2 pronunciation. Most notably, comprehensive work of the mappings between human ratings and machine scores have been intensively researched using extensive speech databases.
