Acoustic Model with Multiple Lexicon Types for Indonesian Speech Recognition

Abstract

Currently, speech recognition datasets are increasingly available freely in various languages. However, speech recognition datasets in the Indonesian language are still challenging to obtain. Consequently, research focusing on speech recognition is challenging to carry out. This research creates Indonesian speech recognition datasets from YouTube channels with subtitles by validating all utterances of downloaded audio to improve the data quality. The quality of the dataset was evaluated using a deep neural network. The time delay neural network (TDNN) was used to build the acoustic model by applying the alignment data from the Gaussian mixture model-hidden Markov model (GMM-HMM). Data augmentation was used to increase the number of validated datasets and enhance the performance of the acoustic model. The results show that the acoustic model built using the validated datasets is better than the unvalidated datasets for all types of lexicons. Utilizing the four lexicon types and increasing the data through augmentation to train the acoustic models can lower the word error rate percentage in the GMM-HMM, TDNN factorization (TDNNF), and CNN-TDNNF-augmented models to 40.85%, 24.96%, and 19.03%, respectively.