Roy Becker's corpus of vowel formants was used as a reference data-set. Vowels with less than 10 entries in the corpus were removed. The formant plot of the train data:
Two separate nearest-neighbour classifiers (with the number of neighbours rather arbitrarily set to 7) were trained for two-formant and three-formant queries. KNeighborsClassifier from the Python package scikit-learn was used.