Triplet Loss-Based Models for COVID-19 Detection from Vocal Sounds

Area of Science:

• Medical acoustics
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• The COVID-19 pandemic necessitated rapid diagnostic tools.
• Vocal biomarkers offer a non-invasive method for disease detection.
• Acoustic analysis of speech and coughs shows potential for identifying respiratory illnesses.

Purpose of the Study:

• To investigate the efficacy of Mel-spectrogram analysis with Convolutional Neural Networks (CNNs) for automatic COVID-19 detection.
• To evaluate the impact of triplet loss functions on improving the separability and robustness of deep learned representations for COVID-19 detection.
• To assess the performance of these models using a dataset of German speakers recorded via smartphones.

Main Methods:

• Utilized Mel-spectrograms as representations of vocal sounds (sustained vowels, coughs, reading speech).
• Employed Convolutional Neural Networks (CNNs) for feature extraction from Mel-spectrograms.
• Implemented a triplet loss function to enhance inter-class representation separability and robustness.
• Conducted experiments on the 'Your Voice Counts' dataset.

Main Results:

• Triplet loss-based models demonstrated suitability for COVID-19 detection from vocal sounds.
• The best performance, an Unweighted Average Recall (UAR) of 66.5%, was achieved using a triplet loss model with speech data from reading.
• The findings suggest that specific vocal tasks can yield more discriminative features for COVID-19 detection.

Conclusions:

• Automatic detection of COVID-19 using vocal sound analysis is feasible.
• Triplet loss functions can improve the performance of deep learning models for this task.
• Speech recorded during reading tasks shows particular promise for COVID-19 detection via acoustic analysis.