4.4.1 Pair-Wise Classifiers for Consonant Categorization

It is used to classify the learner's consonant sound into the place and manner of articulation. If the classifier is trained using native speakers, then the closer the learner is to the training population, the higher the classification should belong to the corresponding category. We use Pair-Wise classifiers that are specifically designed to extract optimal features for every pair of the categories[31], e.g., the classifier for fricative vs. plosive is dedicated to discriminate their differences. It should be pointed out that we perform the classification of categorical articulation instead of consonant recognition itself, because classification of categorical articulation is an easier task and is directly linked to pronunciation instruction. Table 4.1 and Table 4.2 show the open test discrimination rates of the Pair-Wise classifiers trained for this purpose. These results demonstrate that stable discrimination has been obtained for almost all pairs of the categories.

 

samples VLAB ULAB VDEN UDEN POST PALA VVEL UVEL GLOT

VLAB 567 97.1 88.5 99.5 99.2 96.1 93.4 99.6 98.9

ULAB 316 95.8 97.6 97.8 99.2 96.9 97.4 87.8

VDEN 788 98.6 95.3 91.5 85.9 98.1 95.6

UDEN 431 94.9 100 98.9 94.8 98.0

POST 356 99.0 97.3 98.3 94.9

PALA 336 96.5 99.8 98.5

VVEL 230 97.2 97.6

UVEL 229 99.0

GLOT 274

 

 

samples PLOS PLOY NASA NASY FRIC FRIY AFFR TAPF TAPY APPR

PLOS 1159 96.1 94.5 97.5 97.0 98.9 99.5 94.0 99.1 97.9

PLOY 192 99.7 100 98.0 98.9 99.7 98.0 97.7 96.7

NASA 308 98.0 98.9 99.5 100 94.6 99.2 98.0

NASY 144 98.3 98.1 100 98.0 99.1 89.4

FRIC 791 90.8 94.8 97.7 98.6 98.6

FRIY 72 100 99.6 100 99.2

AFFR 187 99.7 100 100

TAPF 154 92.9 97.5

TAPY 72 99.2

APPR 448