Exploring Philippine Mental Health Podcasts: The Efficacy of Latent Dirichlet Allocation and BERTopic for Topic Modeling in Optimizing English-Filipino ASR-Derived Speech Content Analysis

The integration of English and Filipino in speech processing within the Philippines presented unique challenges due to the nation’s linguistic diversity. This study investigated the efficacy of Latent Dirichlet Allocation (LDA) and BERTopic for enhancing topic modeling in Automatic Speech Recognition (ASR)-derived content from Mental Health podcasts. The study aimed to develop and evaluate an English-Filipino Speech Topic Tagger and Sentiment Analysis system to improve comprehension and analysis of multilingual media. Methodologically, the research employed ASR models for transcription, followed by the application of LDA and BERTopic for topic extraction and sentiment analysis. The average word error rate between 21 pairs of true labeled and ASR-derived data was 8.24%. Results indicated that both LDA and BERTopic effectively uncovered thematic structures and sentiments within the podcast content. BERTopic demonstrated superior performance with significantly better coherence scores. These findings contributed to better mental health awareness and reduced stigma in the Philippines, aligning with Sustainable Development Goals (SDGs) 3 and 10. This research offered a framework for future advancements in natural language processing for low-resource languages, aiming to empower underrepresented communities through innovative technological solutions.