Generalization performance of spetro-temporal speech features

Introduction Despite the fact that the dynamic aspects of speech are very important, conventional speech features as Mel Ceptstral Coefficients (Mfccs) [1] and RelAtive SpecTrAl Perceptual Linear Predictive (Rasta-Plp) features [2] capture only stationary spectral information. We could previously show that a combination of conventional speech features with spectro-temporal speech features yields to improved recognition results in noisy speech [3, 4, 5]. We termed those latter features as Hierarchical Spectro-Temporal (Hist) features. They consist of two layers, the first capturing local spectro-temporal variations and the second integrating them into larger receptive fields (compare Fig. 1). This layout was inspired by a recently proposed system for visual object recognition [6]. On the first layer we apply ICA (Independent Component Analysis) and in the second layer we apply different learning algorithms, detailed below. Finally we use a Principal Component Analysis (PCA) to orthogonalize the features and further reduce their dimensionality followed by a Hidden Markov Model (HMM) for the recognition.