In this report we present a Diagnostic tool for ASR systems (DASR). The aim was to develop a tool capable to perform statistical analysis of output of ASR decoding process. Many error patterns in the output might be observable directly by humans, but if number of tracking variables (possible causes of errors) is very high, the task for humans becomes too complex. Machines are able to process as much variables as necessary, and performs a statistical analysis on data as well. We discuss design and implementation of the tool. In addition, we present an example of usage of the tool. This is 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.
DASR: A diagnostic tool for automatic speech recognition
Research report RR-06-182
© EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in Research report RR-06-182 and is available at :
PERMALINK : https://www.eurecom.fr/publication/2123