End-to-End Phoneme Recognition in Bahasa Indonesia with Pretrained Speech Embeddings and 1D-CNN Using CTC

Phoneme recognition in low-resource languages like Bahasa Indonesia is hindered by the limited availability of annotated datasets and linguistic resources. This study proposes an end-to-end phoneme recognition system for Bahasa Indonesia using pre-trained speech embeddings from Wav2Vec2 and a one-dimensional Convolutional Neural Network (1D-CNN) with Connectionist Temporal Classification (CTC). The model is compared against a baseline employing fully connected layers. Our findings highlight the effectiveness of integrating pre-trained speech embeddings and CTC in improving phoneme recognition accuracy. The proposed models utilizing a 1D-CNN kernel size of 9 achieve a phoneme error rate (PER) of 5.27 after 100 epochs and 4.89 after 200 epochs of training, using 156.1 hours of the training dataset.