O-Engineers O-Engineers Aug 2017 | Page 40

accent which leads to misidentification of speaker which ultimately results in performance degrada- tion of recognizer. There are various feature extraction techniques which claims to be robust while performing in such environmental conditions. As far as pre-processing stage is concerned, Linear Predictive analysis (LPC) is assumed to be the first step of feature extraction and other robust cepstral feature derived from lin- ear predictive coefficients are then utilized. Linear Prediction of Speech: Since speech is produced by the movement of ana- tomical structures such as vocal cords, vocal tract, nasal cavity, tongue, and lips. Linear model of speech is first proposed by Fant in late 1950s where the glottal pulse, vocal tract, and radiation are indi- vidually modeled as linear filters. A complete model of speech production is shown in the figure. 40