Benchmark for Speaker Identification using Mel Frequency Cepstral Coefficients on Vowels

Abstract

Aim was to obtain the benchmark for speaker identification using Mel Frequency Cepstral Coefficients (MFCC) on vowels following the nasal continuants in Kannada language. Participants chosen were twenty Kannada speaking male neuro-typical adults, in the age range of 20-30 years. Kannada meaningful words (30) with long vowels /a:/, /i:/, /u:/ following the nasal continuants /m/ and /n/ formed the material. Speech Science Lab Work bench, a Semi-Automatic vocabulary dependent speaker recognition software was used to extract MFCC for the trun- cated (PRAAT) long vowels. Results indicated higher percent correct identification for Condition I (live verse live recording). On comparison among the three vowels following the nasal continuant /m/, /i:/ is better followed by /a:/ and /u:/. Whereas for /n/ the vowel /a:/ is better followed by /i:/ and /u:/. On an average of percentage of correct speaker identification of three vowels compared between the nasal continuant, the vowels following the nasal /n/ (90%) and /m/ (90%) was similar. Condition II (Mobile verse Mobile) and Condition III (Mobile verse Live) was comparatively poorer than Condition I, thus the benchmark was ob- tained. Discussion concludes that during the transmission of voice signals through communication channels, the signals are reproduced with errors caused by distortions from the microphone and channel, and acoustical, electromagnetic interferences and noises affect the transmitting signal. Where speech coding algorithms that are part of Global System for Mobile compress speech signal before transmission, reduce the number of bits in digital representation but at the same time, maintain acceptable quality