A non-invasive speech processing method for analysis of vocal fold cancer based on an iterative EM algorithm

The long term goal of this study is to formulate a speech processing algorithm for analysis and classification of speech pathology due to vocal fold cancer. In this initial phase, a signal processing algorithm is proposed which attempts to quantify the change in production when vocal fold cancer is present. The basic premise is that exact glottal flow estimation is not needed to quantify the change in speech production when a stationary excitation pathology exists. The proposed method is based on maximum likelihood (ML) estimation, which allows for a separation of speech components under healthy and assumed pathology conditions. The method constitutes an iterative approach based on the estimation-maximization (EM) algorithm. An evaluation is performed on speech recordings from vocal fold cancer patients undergoing radiation treatment. An enhanced spectral pathology component (ESPC) results, which is shown to vary consistently between pre and post radiation conditions. It is suggested that general analysis of the ESPC feature can provide a quantitative, non-invasive approach for vocal fold pathology detection and characterization of speech production