Filipino and Bisaya Speech Corpus and Baseline Acoustic Models for Healthcare Chatbot ASR

Philippine public schools often need an efficient and systematic way of assessing children’s physical wellness. With the lack of medical personnel in public schools compounded with the limited financial resources of children’s families, children are often barred from routine checkups. These problems motivated the authors to develop a healthcare chatbot for Filipino children. The chatbot employs an automatic speech recognition (ASR) for its speech-to-text module that offers an alternative way for children to interact with the healthcare chatbot. Currently, there are no readily available ASR systems for conversational Filipino speech of children specifically in the healthcare domain. In this paper, we describe the collection of a Filipino and Bisaya speech corpus for the purpose of developing the ASR module of a chatbot that can monitor the general physical wellness of children. We also present the results of initial experiments on baseline acoustic models for two different phoneme sets for Filipino and Bisaya speech, namely PS27 and PS35. The best-performing acoustic model for Filipino was obtained from PS35 which gave an average word error rate of 4.37%. The best-performing acoustic model for Bisaya was obtained from PS35 which gave an average word error rate of 7.16%.