Customized Acoustic Model using Low-Resource Indonesian Speech Dataset for Short Command Speech Recognition System

The deep neural network technique recently is one of the most state-of-art approaches to train speech recognition systems. In most cases, the neural network needs a huge audio corpus and many speakers to perform well, especially for an acoustic model. One of the examples is the English speech recognition system which has shown accurate performance in many applications. However, the low resource dataset cannot boost performance while using the deep neural network to build an acoustic model. It remains challenging to investigate the usage of hyperparameters in the deep neural network to train the acoustic model. This paper explores proper hyperparameters for low-resource Indonesian audio datasets to gain acceptable performance using time-delay neural networks and convolutional neural networks. The dataset was collected from students using 56 short commands suitable for intelligence assistant applications. In our experiment, the time-delay neural networks can obtain 1.09 % WER. Finally, the best performance is achieved by stacking convolutional neural networks layer after input layer and tuning hyperparameters.