English-Filipino Speech Topic Tagger Using Automatic Speech Recognition Modeling and Topic Modeling

We present an English-Filipino Speech Topic Tagger that transcribes English-Filipino speech audio into text and produces relevant keywords from such audio. The tagger was implemented in two parts by transcribing speech data to text using a Filipino fine-tuned English XLSR-Wav2Vec2 Automatic Speech Recognition (ASR) model then extracting context from the transcription using a generative statistical model used for Topic Modeling, Latent Dirichlet Allocation (LDA). The trained English-Filipino ASR model shows a 26.8% Word Error Rate in the validation set. The Speech Topic Tagger was evaluated through an observation-based approach using different YouTube videos as input and achieved the highest evaluation of 46.15% accuracy and lowest evaluation of 14.28% accuracy in producing relevant top-weighted words.