Electronic International Standard Serial Number (EISSN)
Confusion matrices have been used as a tool for the analysis of speech perception or human speech recognition (HSR) for decades. However, they are rarely employed in automatic speech recognition (ASR) mainly due to the lack of a systematic procedure for their exploration. The generalization of formal concept analysis employed in this paper provides a conceptual interpretation of confusion matrices that enables the analysis of the structure of confusions for both human and machine performances. Generalized formal concept analysis transforms confusion matrices into ordered lattices of confusion events, supporting classic results in HSR that identify a hierarchy of virtual articulatory-acoustic channels. Translating this technique into ASR, a detailed map of the relationships among the speech units employed in the system can be traced to make different sources of confusions apparent: the influence of the lexicon, segmentation errors, dialectal variations or limitations of the feature extraction procedures, among others.