Word-Level Acoustic Modeling of Speech for COVID-19 Detection

Speech-based Artificial Intelligence (AI) approaches have the potential to allow for cost-effective automatic detection of diseases such as COVID-19. Despite promising results in existing literature, many approaches are limited by a disregard for phonetic context, which is vital for understanding the articulatory movements that reveal disease-relevant information. Nevertheless, a few studies have investigated the contributions different phonemes play in the automatic detection of diseases such as Parkinson's Disease, asthma, and COVID-19. Some of these studies have reported superior speech-based detection of diseases when sequencing multiple phonemes together. However, these studies sequenced unusual groups of phonemes to create nonce words, limiting the real-world applicability and utility of these findings. In this paper we propose a novel word-level acoustic modeling approach for speech-based COVID-19 detection that uses natural phoneme sequences and outperforms all other speech analysis models employed on the Cambridge Task 2 Dataset by nearly 10% with an AUC=0.66. Also, we compare the performance of the proposed approach against phoneme and file-level acoustic modeling approaches for COVID-19 detection and demonstrate word-level acoustic modeling's superior classification performance. The authors believe the proposed word-level acoustic modeling approach provides a computationally efficient, interpretable, and flexible representation that can be applied to improve speech-based disease detection tasks for other diseases as well.