Modeling pronunciation variations for non-native speech recognition of Korean produced by Chinese learners

Recognition accuracy for non-native speech is often too low to make practical use of ASR technology in interfaces such as CAPT systems. This paper describes how we adapted Korean ASR system to Chinese speakers for building a Korean CAPT system for L1 Mandarin Chinese learners by modeling pronunciation variations frequently produced by Chinese learners. Based on pronunciation variation rules describing substitutions, insertions, and deletions together with phonological knowledge rules realized in different phonemic contexts, the probability of occurrence of each rule is calculated. These rules are used to generate extended pronunciation lexicon. For each learner level, ASR experiment is conducted, where 21.2% relative WER reduction is obtained. This verifies that variation analysis is useful for modeling Korean produced by Chinese learners.