A Comparison of Decision Tree Classifiers for Automatic Diagnosis of Speech Recognition Errors

Abstract

Present speech recognition systems are becoming more complex due to technology advances, optimizations and special requirements such as small computation and memory footprints. Proper handling of system failures can be seen as a kind of fault diagnosis. Motivated by the success of decision tree diagnosis in other scientific fields and by their successful application in speech recognition in the last decade, we contribute to the topic mainly in terms of comparison of different types of decision trees. Five styles are examined: CART (testing three different splitting criteria), C4.5, and then Minimum Message Length (MML), strict MML and Bayesian styles decision trees. We apply these techniques to data of computer speech recognition fed by intrinsically variable speech. We conclude that for this task, CART technique outperforms C4.5 in terms of better classification for ASR failures.