Phonics Recognition for Indonesian Dialects Using PNCC and RNN-GRU

Phonics plays an important role in helping learners develop reading and spelling skills by linking sounds to letters. However, applying phonics in multilingual environments such as Indonesia can be challenging due to the influence of regional dialects, especially for non-native English speakers. In this study, a phonics recognition system is developed using several deep learning approaches. The dataset consists of 986 audio recordings collected from 38 speakers, including both native and non-native English speakers from different regions in Indonesia. To improve data diversity, augmentation techniques such as pitch shifting and speed perturbation are applied. Feature extraction is performed using MFCC and PNCC, followed by classification using CNN, RNN-GRU, and Transformer models. The results show that the RNN-GRU model with PNCC features achieves the best performance, with an accuracy of 94.59% and an F1 Score of 0.946. Compared to previous work using SVM and MFCC, this approach provides better results. It is also observed that PNCC is more robust in handling pronunciation variations, and that dialect differences have a noticeable impact on model performance. Overall, this study highlights the importance of considering dialect variation in phonics recognition and shows how deep learning can be used to build more adaptive speech-based learning systems.