Abstracts Submitted: 459
Number of Users: 527
View Abstracts Submitted
Back to home Page
Parkinson disease is a chronic neurological disorder characterized by the progressive loss or damage of the dopamine producing cells in the substantia nigra part of the brain. This deficiency affects the brain neuro mediator systems and is responsible for a wide variety of motor and non-motor deficits like falls, imbalances, handwriting changes, mood changes, behavior changes and also speech changes. It is observed that people suffering from this disease suffer speech impairments related to all the subsystems of speech production namely respiration, phonation and articulation. So, speech signal of the diseased might then provide a necessary cue to detect the disease. Also, the acquisition of speech signals being non-invasive can prevent the trauma of patients going through scans and tests for the detection of the disease. In this paper, the most widely used cepstral feature the Mel Frequency Cepstral Coefficients (MFCC) extracted using Sinusoidal Weighted cepstrum Estimation (SWE) technique to detect Parkinson disease in their early stages is proposed. The MFCC coefficients are usually extracted using a single smooth window like the Hamming or Hanning window. By using a single window, the data is not weighted uniformly. Hence to weigh the data more evenly, multiple windowing technique using a set of orthogonal sinusoidal windows is adopted. The sub spectrum obtained using each of these sinusoidal windows is then weighted and averaged to obtain the final spectrum called the sinusoidal weighted spectrum. This final spectrum which is characterized by reduced variance is then used to compute the MFCC coefficients. The MFCC features extracted by single smooth window and a set of sinusoidal windows from the speech samples of healthy people and Parkinson diseased people are then classified using a feed forward neural network classifier. The speech corpus used includes 90 speech samples of healthy people and 98 speech samples of people suffering from Parkinson disease for less than four years. A performance comparison in terms of recognition accuracy, Equal Error Rate (EER), sensitivity and specificity using the two techniques is reported. The results obtained show a maximum improvement in recognition accuracy by 6.4 %, EER by 5.5%, sensitivity by 4.6 % and specificity by 8.2 % using the sinusoidal windowing technique for nine number of tapers in comparison with the single smooth windowing technique.
© Copyright 2017 All Rights Reserved