Benchmark for Speaker Identification Using Long Term Average Spectrum in Kannada Speaking Individuals

Abstract

The identification of people by their voices is a common practice in everyday life. In the last four decades, speaker recognition research has advanced a lot. The aim of the study was to generate bench marking for speaker identification using Long Term Average Spectrum of speech in Kannada speaking individuals. Ten female Kannada speaking normal subjects in the age range of 18-25 years participated in the study. Material included two standard sentences in Kannada developed such that it embedded most of the phonemes in Kannada. Subjects were informed to speak the sentence in a normal modal voice. Samples were analyzed using Long Term Average Spectrum (LTAS) of speech of Computerized Speech Lab (CSL). From the LTAS, kurtosis and skewness were extracted and noted for each speaker. The results revealed several interesting points. Skewness and kurtosis appear to be robust when the number of subjects is limited Its efficiency drops when the number of subjects increased However, a 90% bench marking was obtained for a group of 5 speakers. The results of the present study are restricted to female speakers and Kannada language. Hence generalization of the results to other languages and gender is questionable. Future studies with five speakers in other Indian languages, indirect or mobile recording and disguise conditions are warranted.