Diagnostics of speech recognition : on evaluating feature set performance

In this paper we present an explorative study of diagnostics of speech recognition for finding subsets of features that are most informative in terms of incorrect speech recognition, if variable speech is recognized. The impact on both MFCC and PLP features is investigated. Standard HMM-GMM phoneme-based ASR system with no grammar is used for collection of the all the correct and wrong decodings, and decision tree analysis is used with questions about variance of feature coefficients in the tree nodes. The paper presents various results on importance of quefrency regions in terms of intrinsic speech variabilities, and contributes to better understanding of efficiency of used front-end.