Sinusoidal model-based diagnosis of the common cold from the speech signal

Abstract

Background and objective: Developing speech signal-based non-invasive diagnosis techniques is an emerging research field in biomedical signal processing. Detecting the common cold and other illnesses with similar symptoms may be important to prevent the spread of these viral infections and remotely monitor patient health. This study aims to develop a method that achieves higher performance with fewer features to identify a subject with a common cold from their speech. Method: The spectrum of cold and non-cold speech is distinct from one another. To capture these spectral variations, this study proposed three features: normalized harmonic peak with respect to the first harmonic peak (NHPF), normalized harmonic peak with respect to the maximum value of harmonic peak (NHPM), and successive harmonic peak ratio (SHPR). The NHPF and NHPM feature gives information about the relative magnitudes of all harmonics with respect to the first and maximum harmonics, respectively. The SHPR feature gives information on the relative harmonic magnitude with respect to the magnitude of its neighboring harmonic. The classifiers derive frame-level confidence scores for cold and non-cold classes. For utterance level categorization, confidence scores of all frames are summed, and a class with a higher confidence score is assigned to that utterance. Results: The analysis shows that the proposed features effectively classify cold and non-cold speech. We achieved 69.16% and 66.90% UAR on the develop and test set of the URTIC database. Conclusion: The proposed features efficiently capture the spectral difference between cold and non-cold speech and can be utilized to automatically diagnose common cold and related disorders.