Building Corpora of Spoken Filipino Words Using Speech Segmentation with Automatic Labeling

Abstract

There are hardly any open access speech corpora in Filipino that are structured and can be used to train speech recognition systems that utilize deep learning models. The existing data sets of audios in Filipino language are in forms that require pre-processing and cleansing. This paper proposed a method that would allow building up of Filipino words corpora which can be used as data set for deep learning speech recognition systems. The method utilized speech segmentation technique to extract words from a sound file composed of sentences, words, and syllables. Three trials with variances on the length of silence were conducted to increase accuracy of segmentation and create a more robust Filipino words corpus. Using several Python tools combined with Google Cloud Speech Recognition Application Program Interface, automatic speech segmentation with automatic labeling for Filipino language was achieved with 95% accuracy.