Effectiveness of LP Derived Features and DCTC in Twins Identification - Iterative Speaker Clustering Approach

The main objective of this paper is to explore the effectiveness of LP derived features and DCTC for the purpose of twins identification based on their speech samples. We propose features such as LPCC, LSF and DCTC in evaluating the performance of the system. The features LPCC and LSF are derived by performing LP analysis on speech segments of 16 msecs duration. For extracting DCTC, speech samples are transformed into DCT coefficients first and IDCT is performed on logarithm of DCT coefficients. These features are captured and quantized into M clusters representing L feature vectors by means of K-means clustering approach. Twins identification is done on the basis of finding distance between cluster centroids and feature vectors of the noisy test speech. Speaker is identified based on the minimum average distance between speaker models and feature vectors of noisy test speech. The proposed features are analyzed in this work and experimental results reveal the good performance of the system in terms of sub optimal and true success rates and also perform the comparative analysis between the proposed features.