On the use of cepstral coefficients, multilayer perceptron networks and Gaussian mixture models for Vocal Fold Edema diagnosis

This paper presents further developments towards automatic Laryngeal diseases classification by employing Cepstral coefficients to represent the voice signals, and a classification pipeline employing Multilayer Perceptron Neural Networks and Gaussian Mixture Models for discrimination among normal voice, voices affected by local fold Edema and voices affected by other pathologies (nodules, cysts and paralysis). Laryngeal diseases affect many professionals who use their voices as the main working tool, and conventional diagnosis techniques of these diseases are typically invasive, causing much discomfort to the patient. In recent years, Digital Voice Processing techniques have been investigated to produce non-invasive systems to aid diagnosis by a specialist. This work proposes a method of analysis that is aligned with the non-invasive approach. An experimental evaluation has highlighted the promising results of the proposed method: (a) discrimination between normal and pathological voices with a correct classification rate above 93 % for normal voice and above 94 % for pathological voice; and (b) discrimination between edema and other pathological voices with a correct classification rate above 76% for edema voice and above 85% for other pathologies.