TDNN-HMM ASR Systems on Under-Resourced Local Languages Towards Application in a Healthcare Chatbot

Research in Automatic Speech Recognition (ASR) for under-resourced local languages faces resource limitations. Although there have been previous studies on Filipino ASR, they primarily focus on the Hidden Markov Model (HMM) with the Gaussian Mixture Model (GMM) approach. Studies on Bisaya ASR are much more limited in terms of resources such as speech corpus and previous works. Because of this, there is a lack of neural networks or end-to-end system studies since neural networks require massive amounts of data to train. An alternative to this would be the hybrid model which uses both neural networks and HMM. This neural network architecture would still need data but not as much as an end-to-end ASR system. To address these opportunities, the study used the healthcare chatbot speech corpora in Filipino and Bisaya languages. Furthermore, in this study, we collected, preprocessed and transcribed additional Filipino speech data that were augmented to the existing corpus. Using the aforementioned speech corpora, we developed an HMM-GMM ASR system similar to previous studies as a baseline. In this study, several experiments were performed with various phoneme sets, n-grams, language model weights, HMM states, and model enhancement techniques. The authors found that the best models for Filipino and Bisaya used speaker adaptive training (SAT) with respective word error rates (WERs) of 3.96% and 5.41%. This paper also presents the development of a deep neural network (DNN) HMM baseline model and time delay neural network (TDNN) HMM models with symmetric, asymmetric, and subsampled time strides. For Filipino, the best model is the asymmetric TDNN-HMM model with a 3.48% WER. For Bisaya, the best model is the baseline DNN-HMM model with a 5.50% WER.