Deriving perceptual gradation OF L2 English mispronunciations using crowdsourcing and the WorkerRank algorithm

Pedagogically, feedback in CAPT systems can be improved by focusing on the most critical errors rather than presenting all errors to the users at the same time. This paper presents our work on the use of crowdsourcing for collection of gradations of word-level mispronunciations in non-native English speech. Quality control procedures based on the proposed WorkerRank algorithm (adapted from well-known PageRank algorithm), are performed for selecting a subset of the crowdsourced data in order to ensure reliability. Based on the selected data, we derive a set of rated word-level mispronunciations, according to a four-point gradation of no error, subtle, medium and salient errors.