Speech-based gender recognition using linear prediction and mel-frequency cepstral coefficients

Abstract

Gender discrimination and awareness are essentially practiced in social, education, workplace, and economic sectors across the globe. A person manifests this attribute naturally in gait, body gesture, facial, including speech. For that reason, automatic gender recognition (AGR) has become an interesting sub-topic in speech recognition systems that can be found in many speech technology applications. However, retrieving salient gender-related information from a speech signal is a challenging problem since speech contains abundant information apart from gender. The paper intends to compare the performance of human vocal tract-based model i.e., linear prediction coefficients (LPC) and human auditory-based model i.e., Mel-frequency cepstral coefficients (MFCC) which are popularly used in other speech recognition tasks by experimentation of optimal feature parameters and classifier’s parameters. The audio data used in this study was obtained from 93 speakers uttering selected words with different vowels. The two feature vectors were tested using two classification algorithms namely, discriminant analysis (DA) and artificial neural network (ANN). Although the experimental results were promising using both feature parameters, the best overall accuracy rate of 97.07% was recorded using MFCC-ANN techniques with almost equal performance for male and female classes.

Comparative Analysis of LPC and MFCC for Male Speaker Recognition in Text-Independent Context

“…In speech recognition systems, MFCC is a popular feature extraction approach. MFCC was built on a collection of filter banks consisting of multiple band pass filters in the form of triangle shape window functions that used mel-scale warped frequency to decode speech sounds [20]. Equation ( 5) states the mapping of acoustic linear frequency, 𝑓 into perceptual mel frequency.…”

“…The process was repeated until the final subset was designated as the testing dataset. The results were evaluated using a confusion matrix [16,20].…”