Laryngeal cancer discrimination using linear predictive features

摘要

In this work, we focus on development of non-invasive, cost-effective and user friendly method for automatic discrimination of Laryngeal cancer, an uncontrolled cell growth near human voice box. Proposed system models the significant vocal characteristics for cancer patient discrimination using just their speech as an input to the system. The four distinct mathematical representations of state of the art linear predictive features namely reflection coefficients, log area ratio, Line Spectral Frequency (LSF) and perceptual linear prediction are analyzed. The state of the art linear discriminant analysis is used to segregate probable cancerous voice samples. The dominance of a particular feature is understood using objective performance measures like accuracy, sensitivity, precision, specificity, F-ratio and miss rate. LSF performs slightly better than its counter representations with 89.03% accuracy and miss rate of 5.55%.