Classification of nonverbal human produced audio events: A pilot study

Abstract

The accurate classification of nonverbal human producedaudio events opens the door to numerous applications beyondhealth monitoring. Voluntary events, such as tongue clickingand teeth chattering, may lead to a novel way of silent interfacecommand. Involuntary events, such as coughing and clearingthe throat, may advance the current state-of-the-art in hearinghealth research. The challenge of such applications is the bal-ance between the processing capabilities of a small intra-auraldevice and the accuracy of classification. In this pilot study,10 nonverbal audio events are captured inside the ear canalblocked by an intra-aural device. The performance of three clas-sifiers is investigated: Gaussian Mixture Model (GMM), Sup-port Vector Machine and Multi-Layer Perceptron. Each classi-fier is trained using three different feature vector structures con-structed using the mel-frequency cepstral (MFCC) coefficientsand their derivatives. Fusion of the MFCCs with the auditory-inspired amplitude modulation features (AAMF) is also investi-gated. Classification is compared between binaural and monau-ral training sets as well as for noisy and clean conditions. Thehighest accuracy is achieved at 75.45% using the GMM classi-fier with the binaural MFCC+AAMF clean training set. Accu-racy of 73.47% is achieved by training and testing the classifierwith the binaural clean and noisy dataset.