Robust Feature Extraction based on Spectral Peaks of Group Delay and Autocorrelation Function and Phase Domain Analysis

This paper presents a new robust feature set for noisy speech recognition in phase domain along with spectral peaks obtained from group delay and autocorrelation functions. The group delay domain is appropriate for formant tracking and autocorrelation domain is well-known for its pole preserving and noise separation properties. In this paper, we report on appending spectral peaks obtained in either group delay or autocorrelation domains to the feature vectors extracted originally in phase domain to create a new feature set. We tested our features on the Aurora 2 noisy isolated-word task and found that it led to improvements over other group delay-based and autocorrelation-based methods that use magnitude instead of phase for feature extraction.