Automatic spoken English test for Chinese learners

In this paper, we present research towards developing an automatic spoken language test for Chinese learners of English-as-a-foreign-language. The system can record and process learners´ speech and instantly provide segmental and prosodic scores on pronunciation quality. To achieve high correlation between human and automatic scores, we employed a hidden Markov model based speech recognizer to align speech with the expected answer. Afterwards, segmental feature was assessed as a confidence score at the phoneme level. The prosodic aspects of the pronunciation were examined by using the detected duration, energy and pitch cues. Experiments were performed on a Mandarin Chinese accent speech corpus which was designed and collected by the authors. Results show that these automatic scores enabled reliable estimation of human scores. In particular, we achieved a 0.90 correlation with human scores for the segmental evaluation at the speaker level.