Closed-set speaker identification system based on MFCC and PNCC features combination with different fusion strategies

In this chapter, we investigate novel fusion strategies for text-independent speaker identification. In this context, we present four main simulations for speaker identification accuracy (SIA) including different fusion strategies such as feature-based early fusion, score-based late fusion, early-late fusion (combination of feature- and score-based), late fusion for concatenated features, and statistically independent normalized scores fusion for all the previous scores. The Gaussian mixture model (GMM) with the universal background model (UBM) is used for voice modeling. For the original clean speech recording, Mel frequency cepstral coefficient (MFCC) features are utilized. power normalized cepstral coefficient (PNCC) features are employed in the noisy environment due to their efficiency with noise. Hence, to produce a robust speaker identification system, the MFCCs and PNCCs are combined. In addition, cepstral mean and variance normalization (CMVN) and feature warping (FW) are used in order to mitigate possible linear channel effects, as they are robust for channel and handset mismatch and additive noise. We used the TIMIT database to evaluate the closed-set speaker identification. Results show the highest SIA (95%) at a mixture size of 512 using the late fusion approach. Moreover, a fusion-based technique for MFCC and PNCC features results in better accuracy than using each feature alone.