ROBUST FEATURES FOR NOISY TEXT-INDEPENDENT SPEAKER IDENTIFICATION USING GFCC ALGORITHM COMBINED TO VAD AND CMNTECHNIQUES
El Bachir Tazi, Abderrahim Benabbou, M. Harti · 2012
A major problem of most speaker identification syst ems is their unsatisfactory robustness in noisy environments. The performance of automatic speaker identification systems degrade drastically in the presence of noise and other distortions, especially when there is a noise level mismatch between the training and testing environments. In this experime ntal research we have studied a recently robust fro nt-end algorithm based on Gammatone Frequency Cepstral Coefficients GFCC associated to Voice Activity DetectorVAD and Cepstral Mean Normalization CMN techniques. Our system using a Gaussian Mixture Models GMM classifier are implemented under MATLAB®7 programming environment. An Expectation Maximization EM algorithm was used to maximize the sum of Gaussian densitiesuntil convergence was reached.Evaluation is carried out on our owndatabas e containing 51 mixed Arabic speakers. All test utterances are corrupted by a multilevel White Gaus sian Noise WGN.Our aim is to study the performances of the suggested architecture and m ake a comparison with the conventional Mel Frequency Cepstral Coefficients MFCC method which we have successfully implemented and tested in the previous work. The obtained experimental results co nfirm the superior performance of the proposed method over MFCC and outperform it in different noisy environments. Theevaluationresultsbased on the recognition rate accuracy show that both MFCC and the proposed features extractor have perfects performances in low-noise environments when Signal per Noise Ratio SNR is greater than 35 dB (practically 100% in all cases). But when the SNR o f test signal changed from 0 to 40 dB, the average accuracy of the MFCCs methods is only 52.14%, while the proposed GFCCsfeatures extractors associated to VAD and CMN techniques still achieves an average accuracy of 57.22%.