Spectro - Temporal Based Feature Extraction for Identification of Pathological Voice Signals from Normal

Vocal pathology commonly referred to as vocal disorders or dysphonia that focuses on assessing, diagnosing, and treating diseases that have an impact on the voice's quality, function and output. It entails the evaluation and treatment of a wide range of vocal anomalies and difficulties that may result from functional, neurological, psychological, or physical reasons. One-fifth of the world's population has experienced vocal abnormalities brought on by illnesses or other dysfunctions. Voice disorders are any irregularities in a person's typical speech, which has an impact on social integration and communication. Prior to substantial study, a variety of conventional (invasive) procedures for voice disorder detection were employed. However, computer-based (non-invasive) techniques for voice problem detection are now available. Identification of voice pathology in most of the proposed systems used precise database and obtained lesser accuracy. This work utilizes various number of voice signals for the vowel /a/, /i/ and /u/ for both healthy and pathology from Saarbrucken voice database. Various features such as MFCC, MFCC delta, pitch, crest, entropy, flatness, flux, kurtosis, roll off point, skewness, slope, spread, harmonic ratio and spectral centroid were extracted for both healthy and pathological voices. Feature selection methods such as principal component analysis and fuzzy entropy measures were used to obtain prominent features. MFCC, MFCC delta and spectral centroid were found as a significant feature from FEM which helps in categorization of various voice disorders.