Improving the Noise Robustness of Prominence Detection for Children's Oral Reading Assessment

Abstract

Reading skill is a critical component of basic literacy. We aim to develop an automated system to assess oral reading skills of primary school children (learning English as a second language) that could eventually be valuable in the scenario of teacher shortage typical of rural areas in the country. This work focuses on the rating of prosody, an important aspect of fluency in speech delivery. In particular, a system for the detection of word prominence based on prosodic features is presented and tested on real-world data marked by background noise typical of the school setting. To counteract the observed drop in prominence classification accuracy, two distinct approaches to noisy speech enhancement are evaluated for various types of background noise. A recently proposed Generative Adversarial Network(GAN) based method is found to be effective in achieving noise suppression with low levels of speech distortion that minimally impact prosodic feature extraction. The implementation and training of the GAN system is discussed and insights are provided on its performance with reference to that of classical spectral subtraction based enhancement.