Effect of Noise Reduction Technique on Speaker Identification Using Mel-Frequency Cepstral Co-Efficients of Long Vowels

Abstract

Speech is always accompanied by noise when the speaker is talking in the environment. To improve the intelligibility of speech signal, noise should be reduced using noise reduction softwares. From the existing software the aim of the present study was to examine the eect of noise reduc- tion technique on speaker identication using Mel Frequency Cepstral Co-Ecient (MFCC) on the long vowels in Kannada language. Ten Kannada speaking neuro-typical adults in the age range of 20-35 years (5 males and 5 females) participated in the study. Commonly occurring Kannada meaningful sentences with long vowels /a:/, /i:/, /u:/ was used for reading task. The same was recorded in two dierent conditions: Lab condition and Trac condition. These samples were analyzed under two phases: Before noise reduction (BNR) and After noise reduction (ANR), using Sound Cleaner Software. Speech Science Lab Work bench software was used to extract MFCC for the truncated (PRAAT software) vowels. Results of the study revealed that in Lab condition, Trac condition (BNR), Trac condition (ANR), Lab condition verses trac (BNR) and in Lab condition verses trac condition (ANR), the vowel /i:/ is found to be better followed by /a:/ and /u:/ in the average percentage of correct speaker identication of the vowels. Overall results revealed vowel /i:/ is better for speaker identication. Hence, the `sound cleaner' has a signicant eect on percent speaker identication by reducing the in uence of noise without majorly aecting the acoustical parameter of certain vowel considered for the present study.