The Use of SDAE in Noisy English Mispronunciation Detection and Diagnosis towards Application in Mobile Learning
Yizhi Wu, Junjie Zhang, Quan Dong · 2019
The problem of noise in the environment is a bottleneck in the application of computer assisted pronunciation training (CAPT) system. In this paper, in order to solve the noise problem in mobile learning, a mispronunciation detection and diagnosis (MD-D) model based on stack denoising autoencoder (SDAE) is proposed for noisy English phonemes. We built a SDAE based recognition network to detect mispronunciation in noisy speech. SDAE is used to obtain robust data features in different noisy environments. The classifier layer at the top of the network is used to generate posterior probability of error types. The parameters of the SDAE network are obtained through unsupervised pretraining and supervised fine-tuning. We used rectified linear unit (ReLU) as the activation function of SDAE to get more advanced features. The verification experiment used SDAE-based MDD model to obtain a better mispronunciation detection performance using the dataset of NOISE-92 and the Speech Accent Archive pronunciation corpus with varied SNR than the performance of SVM. The experimental results showed that our work provided a new example in mispronunciation detection under the situation of noisy mobile learning.