Go To Program 54 On the Statistical Properties of Reverberant Speech Feature Vector Sequences Armin Sehr 1 Walter Kellermann 1 1 Multimedia Communications and Signal Processing, University of Erlangen-Nuremberg, Erlangen, Germany The statistical properties of reverberant logarithmic mel-spectral feature vector sequences that are relevant for acoustic modeling in robust speech recognition are analyzed and compared to the corresponding properties of clean-speech feature vector sequences. The investigation focuses on probability densities of feature vector elements and the correlation between feature vectors of different frames. A Monte-Carlo method is used for a quantitative analysis of the density changes and the increase in inter-frame correlation due to reverberation. As an example for the insights that can be obtained from this analysis, the densities and inter-frame correlations of reverberant features are compared to the modeling capabilities of reverberantly-trained and adapted HMMs. Thus, limitations of these approaches can be clearly identified. View pdf