Extension of the remos concept to frequency-filtering-based features for reverberation-robust Speech recognition

The introduction of partly decorrelated features into the REMOS (REverberationMOdeling for Speech recognition) concept for distant-talking speech recognition [1] is discussed. REMOS combines a hidden Markov model (HMM), trained on clean speech, with a reverberation model capturing certain room characteristics. The most likely contributions of both models to a reverberant observation are determined by an inner optimization problem. In HMM frameworks, decorrelated features are assumed when diagonal covariance matrices are used in the output densities. However, in REMOS, only highly correlated logmelspec (logarithmic mel-spectral) features have been used so far, which has been limiting the recognition performance. In this work, we extend the RE-MOS concept and introduce a new set of partly decorrelated features derived from the frequency filtering [2]. Recognition experiments with connected digits show a consistent relative reduction in word error rate of up to 29% compared to the former logmelspec implementation.