Identification of convict by using several biometrics namely face, finger print, iris, DNA patterns and so on is currently in vogue. Speech is seldom used for such tasks since this process becomes more challenging if it is needed to identify an offender from among identical twin pairs because twins exhibit similar characteristics in contemplating manners, style and expression of emotions. In this work, cost effective speech is used as a biometric to identify twin among twin pairs including identical twins. This work involves the use of perceptual features and deep learning techniques to identify a twin and the performance of the system is analysed by using recognition accuracy, false rejection rate and false acceptance rate as metrics. Deep machine learning techniques namely Recurrent neural networks and Convolutional neural networks are used for creating models for twin pairs. Among the perceptual features, features with filters spaced in Equivalent Rectangular Bandwidth (ERB) and MEL scale have provided 100% as accuracy for both text dependent and text independent modes of classifying the respective twin among the twin pairs based on Convolutional neural network. Perceptual features with filters spaced in BARK scale have provided better accuracy of 98% as compared to other perceptual features for text dependent and independent twin identification by using RNN based machine learning classifier. Hence the experimental results explicitly shows that, proposed method works effectively for twins identification compared to the other existing methods.