G2P and ASR techniques for low-resource phonetic transcription of Tagalog, Cebuano, and Hiligaynon

Philippine linguists are tasked with documenting over 170 indigenous languages. A key part of this documentation is the phonetic transcription of recorded speech, which is typically done by hand, and is often expensive and time-consuming. Automated phonetic transcription systems provide a faster and cheaper alternative to manual transcription, but no such system has yet been developed for most Philippine languages. In this paper, we present an implementation of three APT methods—grapheme-to-phoneme conversion, automatic speech recognition, and adaptive alignment—for transcription of small speech corpora in Tagalog, Cebuano, and Hiligaynon. We show that the G2P, adaptive, and select ASR models perform at par with human transcribers while greatly reducing total time and costs. These systems serve as a competent baseline for future developments in APT for Philippine languages, and are expected to facilitate further research and advancements in Philippine linguistics and speech technology.