Channel Robust MFCCs for Continuous Speech Speaker Recognition

Abstract

Over the years, MFCC (Mel Frequency Cepstral Coefficients), has been used as a standard acoustic feature set for speech and speaker recognition. The models derived from these features gives optimum performance in terms of recognition of speakers for the same training and testing conditions. But mismatch between training and testing conditions and type of channel used for creating speaker model, drastically drops the performance of speaker recognition system. In this experimental research, the performance of MFCCs for closed-set text independent speaker recognition is studied under different training and testing conditions. Magnitude spectral subtraction is used to estimate magnitude spectrum of clean speech from additive noise magnitude. The mel-warped cepstral coefficients are then normalized by taking their mean, referred as cepstral mean normalization used to reduce the effect of convolution noise created due to change in channel between training and testing. The performance of this modified MFCCs, have been tested using Multi-speaker continuous (Hindi) speech database (By Department of Information Technology, Government of India). Use of improved MFCC as compared to conventional MFCC perk up the speaker recognition performance drastically.