SVM-Based Detection of Misannotated Words in Read Speech Corpora

Automatic detection of misannotated words in single-speaker read-speech corpora is investigated in this paper. Support vector machine (SVM) classifier was proposed to detect the misannotated words. Its performance was evaluated with respect to various word-level feature sets. The SVM classifier was shown to perform very well with both high precision and recall scores and with F1 measure being almost 88%. This is a statistically significant improvement over a traditionally used outlier-based detection method.