Performance Comparison of EMD based Noise Classification for different SNR using GMM and k-NN Classifiers

In today’s era of digital revolution, electronic systems in the context of audio communication typically perform transmission, playback, analysis and synthesis of audio signals. So, from the perspective of electronic system/product design for any of these purposes, noise influences must be carefully considered. Various types of noise and distortion can be characterized and number of techniques can be assisted in mitigating their effects, thus enhancing the quality and intelligibility of the speech signal. The first and foremost step towards this characterization is noise classification. In this concern, this paper addresses the issue of environmental background noise classification using Empirical Mode Decomposition (EMD). Instead of using an apriori choice of filters or basis functions to separate a frequency component, the EMD typically expands the time series into a set of functions defined by the signal itself; commonly known as Intrinsic Mode Functions (IMFs). These IMFs (and not the actual signal) are then used for feature extraction. This work suggests hybrid feature vectors for classification and proposes an optimized best suitable feature set for classification of different noisy environments with variation in signal-to-noise ratio (SNR) level. For classification, Maximum-Likelihood Gaussian Mixture Model (ML-GMM) and k-Nearest Neighbor (k-NN) classifiers are used. Utilization of this optimized feature set yields the maximum accuracy in multiclass noise classification irrespective of SNR variation. Keywords— Empirical Mode Decomposition, Intrinsic Mode Function, k-Nearest Neighbor classifier, MaximumLikelihood Gaussian Mixture Model.