Investigating Speech Attribute Features for Anti-Phone based Pronunciation Verification Approach
Ayat Hafzalla Ahmed, Hager Morsy, Sherif Mahdi Abdo · International Journal of Computer Applications · 2021
With increased computing power, there has been a renewed interest in computer-assisted pronunciation learning (CAPL) applications in recent years; Automatic accurate pronunciation verification method plays an important role in automating the learning process and increasing its quality.Pronunciation errors can be divided into phonemic and prosodic error types.In this paper we propose a phonemelevel pronunciation verification method for Quranic Arabic based on anti-phone model.For each phoneme a binary support vector machine (SVM) classifier is trained to distinguish each phoneme from other phonemes.The (SVM) classifier is trained using speech attribute features derived from a bank of speech attribute detectors, namely manners and places of articulation.The feed forward deep neural network (DNN) architecture is utilized for the speech attribute detectors.The system is evaluated against speech corpora collected from fluent Quran reciters and achieved phonemelevel false-acceptance and false-rejection rates ranging from 2% to 25%.