A Robust Speaker Identification System for Natural and Whispered Speech

Speaker Identification, using both speech and whisper, is an emerging research topic. The main challenge lies in improving the robustness of the system in highly noisy environment. In this paper, different identification algorithms for both normal and whispered speech have been compared to check the robustness. Mel frequency cepstral coefficient (MFCC) method, Gabor filter-bank methods and an Empirical Mode Decomposition (EMD) based AM-FM approach have been employed for feature extraction. The extracted features have been classified with various classifiers such as Support Vector Machine, Fine K Nearest Neighbor (KNN) and Weighted KNN. A database of 16 subjects has been created, both in normal, as well as in whispered mode. It is observed that the separate Gabor filter-bank method provides the best accuracy (98.1% with Fine KNN) for whispered speech, and the AM-FM approach offers the best accuracy (98.9% with Fine KNN) for normal speech.