Benchmark for Speaker Identification using Mel Frequency Cepstral Coefficients on Vowels Preceding Nasal Continuants in Kannada

Abstract

Identification of speakers in forensic context is generally about comparing voices. In forensic speaker identification, the serious problem is to identify an unfamiliar speaker whose voice has been recorded at some stage in the committing of a crime. Vowels, nasals and fricatives (in decreasing order) are usually suggested for voice recognition because they are somewhat easy to identify in speech signals and their spectra contain features that reliably differentiate speakers based on semi-automatic methods. In this context, the aim of the present study was to obtain the percentage of speaker identification using vowels preceding nasal continuants in Kannada speaking individuals using semi-automatic method. The participants chosen for the study were twenty Kannada speaking adult males in the age range of 21-32 years constituted as Group I. This was further sub grouped (participants reduced) as Group II constituting ten speakers. The material was meaningful mono-, bi-, and/or multisyllabic Kannada words containing long vowels /a:/, /i:/ and /u:/ preceding nasal continuants /m/ and /n/ embedded in Kannada sentences. The participants read the material four times each under two conditions (a) live recording and (b) mobile network recording which were stored into the computer memory. The target words were truncated using the PRAAT software. Each vowel preceding nasal was subjected for Mel Frequency Cepstral Coefficients (MFCCs) using Speech Science lab Workbench for Semi-automatic speaker recognition (vocabulary dependent) software. The same was found across the three conditions when the participants reduced from twenty to ten in number. The study was compared under three conditions: (a) Live vs live recording, (b) Mobile network vs mobile network recording and (c) Live vs mobile network recoding. The results of the present study indicated quite high percent of correct speaker identification using MFCCs in Live vs Live and Mobile network vs Mobile network conditions compared to Live vs mobile network condition. Thus, the present study provided some proof to look at the efficiency of semi-automatic method using MFCC which helps in speaker identification. The obtained outcome would serve as potential measure in the forensic scenario for identification of speakers using vowels preceding nasal continuants in Kannada.