Novel Feature Vector Set Extraction using Spectral Peaks in Autocorrelation Domain

This paper presents a new feature vector set for noisy speech recognition in autocorrelation domain. The autocorrelation domain is well known for its pole preserving and noise separation properties. In this paper we will use the autocorrelation domain as an appropriate candidate for robust feature extraction. In our approach, extraction of mel frequency cepstral coefficients (MFCC) of the speech signals are proposed based on novel Differentiated Relative Higher Order Autocorrelation Coefficient Sequence Spectrum (DRHOASS). In this approach, initially the lower lags of the noisy speech autocorrelation sequence are discarded and then, the effect of noise is further suppressed using a high pass filter in autocorrelation domain. Finally, the feature vector set of the speech signal is found using the spectral peaks of the filtered autocorrelation sequence. We tested our features on the Hindi isolated-word task and found that it led to noticeable improvements over other autocorrelation-based and differential spetral-based methods.