Robust speaker identification system over AWGN channel using improved features extraction and efficient SAD algorithm with prior SNR estimation
Résumé
This paper motivates the combination of Autoregressive (AR) parameters and Mel Frequency Cepstral Coefficients (MFCC) features for remote robust text-independent speaker identification. All speaker identification techniques start by converting the raw speech signal into a sequence of acoustic feature vectors carrying distinct information about the signal. The most commonly used acoustic vectors are Mel Frequency Cepstral Coefficients (MFCC) which is a very useful feature for speaker identification system but it deteriorates in the presence of noise, thus to have better identification rate, we have developed a robust feature extraction method based on the combination of MFCC and Autoregressive model (AR) parameters modeled with GMM model. To improve the identification rate accuracy, an efficient speech activity detection (SAD) algorithm based on prior SNR estimation are proposed in the pre-processing phase. To validate our work TIMIT database with speech from 630 speakers has been used. The first four utterances for each speaker could be defined as the training set while 1 utterance as the test set. The use of AR-MFCC approach has showed significant improvements in identification rate accuracy when compared with MFCC. However, in terms of runtime, AR-MFCC requires more time to execute than MFCC. Our SAD algorithm has provided a suitable contour of speech activity in noisy conditions
Domaines
Informatique [cs]Origine | Fichiers produits par l'(les) auteur(s) |
---|