Investigating Gammatone Filterbank-Based i-Vectors for Speaker Verification Task

Abstract

AbstractMost of the recent studies based on i-vector framework use conventional mel-frequency cepstral coefficient (MFCC) features. However, the main issue that is still faced with MFCC features is its sensitivity to noisy conditions. The verification performance degrades significantly in noisy conditions. In this paper, we replace the most commonly used MFCC features with new features known as gammatone cepstral coefficients (GFCCs) features. These features are built on gammatone filter bank. The speech data used in the experimentation are taken from TIMIT dataset. Our results show that combination of proposed GFCC features and conventional MFCC features in i-vector framework outperforms individual features not only for representing speaker specific information but also for making our verification system robust in noisy environment. Our experimental evaluations show that GFCC features outperforms the MFCC features with an improvement in equal error rate (EER) by 1% in matched noise conditions and 0.45% in mismatched noise conditions.KeywordsMFCCGFCCi-vectorSpeaker verificationSNRTIMIT