Semi-automatic approach to ASR errors categorization in multi-speaker corpora

Error diagnosis is an integral part of improving the quality and robustness of any ASR system, especially for languages with limited resources. This paper explores a semi-automatic approach to error categorization usable for databases that have a set of identical sentences produced by a sufficiently large number of speakers. We use a matrix created from an ordered list of speakers and an ordered list of sentences based on the recognizer performance. An algorithm that searches through the errors using such a matrix is proposed and the utilization of information obtained from the output is discussed.