Developing an Automatic Speech Recognizer For Filipino with English Code-Switching in News Broadcast

Abstract

Closed-captioning systems are well-known for video-based broadcasting companies as society transitions into internet-based information consumption. These captioning systems are utilized to cater to most consumers. However, a captioning system for the Filipino language is not readily available to the public. News anchors in the Philippines tend to incorporate a code-switching behavior that mixes English and Filipino languages, which are the two major languages that Filipinos use. The goal of this research is to develop an automatic speech recognizer (ASR) for a captioning system for Filipino news broadcast domain videos. Experiments on finding the optimal speech models and features, and on how code-switching affects the system were conducted. Best results were obtained by using linear discriminant analysis with maximum likelihood linear transform (LDA+MLLT) and speaker adaptive training (SAT) for acoustic modeling. Initial investigation also shows that there is no general pattern for the ASR's performance as a function of code-switching frequency.