Text independent voice based students attendance system under noisy environment using RASTA-MFCC feature

Abstract

This paper motivates the use of RASTA-MFCC (RelAtive SpecTrA-Mel Frequency Cepstral Coefficients) feature and GMM-UBM modeling for text independent voice based students' attendance system under noisy environment. MFCC has been identified as an efficient feature for identifying the speaker because it extracts speaker specific information. The performance of even best speaker identification system with MFCC feature degrades in uncontrolled communication environment. RASTA processing of speech improves the performance of identification system even in the presence of convolutional and additive noise. This paper combines the best of these two processes to yield RASTA-MFCC feature which is robust to noise and also contributes speaker dependent information to identify the speaker efficiently. GMM-UBM (Gaussian Mixture Model-Universal Background Model) modeling technique is used for its faster training and relatively easier updating of new speakers. Experimental result of 93.2% accuracy for Triangular filter bank and 94.5% accuracy for Gaussian filter bank are obtained for 50 speakers of MEPCO speech database in presence of additive and convolutive noise in the context of voice based students' attendance entry.