Classification of Cold and Non-Cold Speech Using Vowel-Like Region Segments

Abstract

This work uses vowel-like region segments of speech to classify cold and non-cold speech signals. As various articulators are affected by the common cold, speech produced during the common cold gets affected. These changes in a speech during common cold can be used to classify cold and non-cold speech. Vowel-like region (VLR) in speech includes vowels, semi-vowels, and diphthongs phonemes. Vowel-like regions are the dominant part of the speech signal. Hence, we have considered only vowel-like regions for cold and non-cold speech classification. The VLRs are identified by locating the VLR onset point (VLROP) and end point (VLREP). The Hilbert envelope and zero frequency filtering methods are used for detection of VLROPs and VLREPs. Mel frequency cepstral coefficients (MFCCs) feature are extracted from VLRs, and the performance of these features are evaluated using a deep neural network. Features extracted from VLRs give comparable results to features extracted from complete active speech (CAS) signal. Compared to the CAS technique, the number of frames that needs to be processed utilizing VLRs is significantly less.