A Comparative Analysis of Noise Robust Speech Features Extracted from All-Pass Based Warping with MFCC in a Noisy Phoneme Recognition

Abstract

In this paper, we investigate the noise robustness of three features, namely, the warped discrete Fourier transform cepstrum (WDFTC), perceptual minimum variance distortionless response (PMVDR) and Mel-frequency cepstral coefficients (MFCC). Here, WDFTC and PMVDR features are generated by adopting all-pass based warping and for the MFCC, we know that spectral warping is generally employed. The PMVDR and WDFTC use warped-LP and warped discrete Fourier transforms, respectively. Particularly, we employ the WDFTC, PMVDR and MFCC features in a continuous noisy monophone recognition task using the TIMIT corpus and a wide variety of acoustical noise types at different SNRs (signal-to-noise ratios). Further, we test these features on a gender-specific monophone recognition task. Finally, we report the recognition performance and discuss many interesting properties of these features. Our study shows that the PMVDR and WDFTC achieve recognition performance superior to the MFCC in noisy conditions.

Frequency warping and robust speaker verification: a comparison of alternative mel-scale representations

“…1) include conventional MFCCs and an alternative implementation via direct warping, two warped all-pole models [23,24] and a recent histogram-based technique producing sparse features [21]. Even though some of the methods have been studied in both ASR [25] and speaker verification [26,27], we feel that it is time to present a self-contained summary and comparison within a single study, using a modern i-vector system [22]. We hope the reader finds our study a useful summary of methods otherwise scattered across the literature.…”