Non-Speech Sounds Classification for People with Hearing Disabilities

People with hearing disabilities experience the problems that stem from not being able to detect or identify sounds on a daily basis. Studying the techniques and algorithms which enable this task to be performed automatically may lead to significant technological progress which will offer huge benefits to deaf people. With the objective of developing an application which is capable of detecting and classifying the different sounds that may emerge in the home, a study is being carried out which shows the most important parameters for processing impulsive sounds such as door bells, alarm clocks, a baby crying... which obtain high accuracy ratios and give the classifier high reliability. To date, an initial prototype has been developed which implements a GMM (Gaussian Mixture Model) classifier which is based on the Gaussian probability distribution for sound event prediction. In order to check the classifier's accuracy, typical speech recognition parameters have been used, such as MFCC (Mel frequency cepstral coefficient), as well as parameters used to recognise musical instruments and background sounds: Spectral Centroid, Roll-Off Point and ZCR (16 parameters in total). By varying a series of factors (number of parameters, the sounds used to train the classifier...) the GMM's behaviour has been analysed obtaining results with over 90% accuracy in frames and up to 100% accuracy using the sound average, identifying doors, telephones and alarm clocks.