Feature subset selection using a new definition of classifiability
Pattern Recognition Letters
| Hi-index | 0.00 |
This paper presents a word-independent technique for classifying the syllable stress of spoken English words. The proposed technique improves upon the existing word-independent techniques by utilizing the acoustic differences of various syllable nuclei. Syllables with acoustically similar nuclei are grouped together and a separate stress classifier is trained for each such group. The performance of the proposed group-specific classifiers is analyzed as the number of groups is increased and is also compared with an alternative data-driven clustering based approach. The proposed technique improves the syllable-level accuracy by 5.2% and the word-level accuracy by 1.1%. The corresponding improvements using the data-driven clustering based approach are 0.12% and 0.02%, respectively.