Impact of Pre-processing on the Performance of Text-Independent Speaker Recognition System

In this paper, we explore the effects various pre-processing techniques have on the performance of a text-independent speaker recognition system. The methods used in these analysis are Spectral Subtraction Method (SSM), Spectral Subtraction Method using Oversubtraction (SSMO), Wiener Filtering and Minimum Mean-Square-Error Short-Time Spectral Amplitude Estimator (MMSE-STSA), followed by a recently proposed voice-activity-detection (VAD) algorithm. For this analysis, we extract the Mel-frequency cepstral coefficients (MFCCs) features to distinguish and verify the identity of various individuals using two different modelling methods, namely, (i) Gaussian Mixture Model (GMM) [using Expectation-Maximization (EM) algorithm] and (ii) Vector Quantization (VQ) [using Linde-Buzo-Gray algorithm], by mapping and classifying the MFCC features. The focus of this work is not to explore the latest or best performance algorithm for a speaker recognition system, but rather a robust and stable, hardware friendly, real-time system, specifically for internet-of-things (IoT) applications. For this purpose, we have manually collected data recorded by different individuals, using different devices, in different environments, from various locations in India. The results indicate that among the eight models trained using pre-processing techniques with 47 users enrolled, six of them gave a recognition accuracy of more than 95%, with the best accuracy of 97.8%.