Use of perceptual features in iterative clustering based twins identification system

The main objective of this paper is to explore the effectiveness of perceptual features in identifying twins by including the interference from pseudo random noise and background conversation on test speech. An algorithm is developed for identifying twins by extracting the proposed features on speech segments of 16 msecs duration. In this algorithm, these features are captured and quantized into M = L/10 clusters representing L feature vectors of training speech. Twins are identified based on the minimum average distance between speaker models developed on clean speech and noisy test speech vectors. These perceptual features are analyzed in this work and the experimental results reveal the comparative performance of the proposed features under various SNR conditions for the speech database containing speakers in the same age group. The noteworthy feature in this work is the theoretical validation of experimental results and performance evaluation based on the reduction in training and test data.