Experimental study on prosodic error detection in English utterances using DNN-based acoustic models trained with prosodic features and labels

摘要

We aim at automatic detection of inade-quate articulation and/or inadequate control ofprosody in L2 speech segments which are judgedas incomprehensible subjectively by humans orobjectively by machines. Since it is known thatL2 English utterances with incorrect stress as-signment tend to be difficult for native listenersto understand, we focus on utterances with in-correct stress assignment.In our previous studies, we proposeda novel method to detect incomprehensible seg-ments from L2 utterances. Native listeners areasked to shadow given L2 utterances. Smooth-ness of shadowing was found to be highly corre-lated with comprehensibility subjectively ratedby shadowers [1, 2, 3]. However, our proposedmethod was not fully automatic and native lis-teners were always needed, but obtained shad-owing utterances and automatically-calculatedsequences of shadowability scores can be viewedas shadowability annotation of the L2 utter-ances used. With a sufficient size of paired cor-pus of L2 utterances and their correspondingshadowability sequences, it may be possible tobuild a virtual shadower, which takes L2 utter-ances as input and generate their correspondingshadowability sequences as output.