Localized spectro-temporal cepstral analysis of speech

Abstract

Drawing on recent progress in auditory neuroscience, we present a novel speech feature analysis technique based on localized spectrotemporal cepstral analysis of speech. We proceed by extracting localized 2D patches from the spectrogram and project onto a 2D discrete cosine (2D-DCT) basis. For each time frame, a speech feature vector is then formed by concatenating low-order 2D-DCT coefficients from the set of corresponding patches. We argue that our framework has significant advantages over standard onedimensional MFCC features. In particular, we find that our features are more robust to noise, and better capture temporal modulations important for recognizing plosive sounds. We evaluate the performance of the proposed features on a TIMIT classification task in clean, pink, and babble noise conditions, and show that our feature analysis outperforms traditional features based on MFCCs.

Point process models of spectro-temporal modulation events for speech recognition

“…The class of spectro-temporal modulation (STM) features have been recently evaluated for speech applications including speech/non-speech discrimination [11], phonetic recognition [12], keyword spotting [13], and word recognition [14]. However, in each case a frame-based strategy was employed, which ultimately relied on dimensionality reduction to improve computational tractability.…”

Phone recognition experiments with 2D-DCT spectro-temporal features

“…With this step we smooth out the unnecessary fine details from the spectrum and reduce feature dimensionality at the same time. Bouvrie et al proposed keeping only the 6 lowest-order 2D-DCT coefficients corresponding to the lower left 3x3 triangle of the coefficient matrix [9], while here we are going to keep 9 coefficients. The right hand side of Fig.…”

“…These findings motivate those feature extraction methods that process the spectro-temporal representation in patches which are localized in both time and frequency. The most popular among these methods is to analyze these localized timefrequency patches by means of Gabor filters [8], but here we are going to follow the study of Bouvrie et al [9], and apply two-dimensional DCT to process them. Fig.…”

Phoneme recognition using Boosted Binary Features

“…Dynamic information is subsequently added by appending approximate temporal derivatives of the cepstral features. Other features such as TRAPS/HATS [1], frequency domain linear prediction features [2], multiresolution RASTA features [3], 2D-DCT localized features [4] extract information directly from the spectro-temporal plane.…”