Speech systems classification based on frequency of binary word features

This paper presents a robust method to classify the speech communication systems on the basis of speech coding technique used in digitizing the speech. The method works for both clear and noisy speech conditions, when the level of noise may be unknown or known in terms of bit alterations level up to 30% and can identify coding in as short as 640 bit speech. The noise level is estimated using a technique based on higher order statistics (HOS) using digitized speech output and is independent of the coding used. The classification algorithms used are of two types: Minimum Distance Classifier (MDC) based on Linear Discriminant function (LDF) and Artificial Neural Net (ANN). The features used are based on frequency of binary words, characterizing the speech coding technique used in the speech systems. Three types of codings namely Pulse Code Modulation (PCM), Continuously Variable Slope Delta Modulation (CVSD) and Linear Predictive Coding (LPC) are considered. The classification score obtained for both known and unknown noise level arc compared. It is found that results are far better when level of noise is known in both types of classifiers.