The Diagnostic Evaluation of Switchboard-corpus Automatic Speech Recognition Systems

Abstract

To see whether the related mistake  patterns can be linked to a particular set of variables, a  Eight Control equipment recognizing (and six forced alignment) algorithms were evaluated for clinical diagnosis.  Each recognizing service's result was converted to a  standardized way and evaluated to a comparative record  made from pronunciations labelled data (which included 54  minutes of information from several hundred speakers). A  job evaluation was used to relate a combination of acoustic,  morphological, etc. speaker attributes to acknowledgment  occurrences throughout this reference data. The decision  trees show that correct categorization of phonetic segments  and characteristics is one of the most constant variables  linked with better recognition performance. These findings  indicate that enhancing the pronouncing modelling used in  verbal pairing, including the acoustic modeling techniques  utilized for morphological classification, might improve  future-generation recognition systems.