Deep Neural Network for Automatic Speech Recognition from Indonesian Audio using Several Lexicon Types

Recently, automatic speech recognition has benefited from advances in deep neural networks (DNNs) to train and deploy speech recognition models. Speech recognition models enable computers to recognize and translate spoken language into text. In this paper, we present an approach to creating an Indonesian voice-to-text dataset using audio collected from YouTube channels and to evaluating the speech recognition models using several lexicon types. The lexicons are created from unique words of the speech corpus. We compared the performance of Time Delay Neural Network Factorization (TDNNF) and Gaussian Mixture Model - Hidden Markov Model (GMM-HMM) models for mono phone, DELTA+DELTA-DELTA, and Speaker Adaptive Training (SAT) based on the %WER when trained using unvalidated and validated datasets. The results showed that there was no significant difference in the %WER among the lexicon types. Moreover, the results revealed that the models trained using a validated dataset perform better than an unvalidated dataset. Additionally, lexicon enmap_kv_vocab_full returned the best result with 29.41% WER when trained using the TDNNF model on an unvalidated dataset. However, lexicon enmap_vocab_l_char provided the best result, with 11.35% WER, when trained using the TDNNF model on a validated dataset.