Detection and Correction of English Speaking Pronunciation Errors Based on Acoustic Pronunciation Model

Abstract

A pronunciation error detection system with accuracy and timely feedback is an important tool for English-speaking learners to improve their learning. In this paper, firstly, the Hidden Markov Model is used to model the individual phonemes and their sequences in speech and the acoustic feature vectors extracted from the input speech signal. Secondly, the spoken English pronunciation error detection algorithms are established as the posterior probability algorithm and the GOP algorithm, respectively. An English-spoken pronunciation error detection system is constructed based on the pronunciation error detection algorithm. Finally, two different criteria are used to evaluate the system's performance from different perspectives. By comparing the recognized phoneme sequences with the standard pronunciation text for alignment, the learners’ mispronunciation can be diagnosed, and the error information can be fed back to the learners. The system evaluation results show that the system’s recognition results differ from the experts’ judgments by more than 23% when using the full extended pronunciation dictionary. In the comparison results with different experts’ manually annotated texts, a correct rate of more than 85.2% and an accuracy rate of 82.7% were obtained. The trimmed extended pronunciation dictionary can detect the wrong pronunciation more effectively. The spoken English pronunciation checker system can accurately identify and provide timely feedback on learners’ incorrect pronunciation.