State dependent feature component selection for noise robust ASR

The acoustic environment in which speech is recorded has a strong influence on the statistical distributions of observed acoustic features. In order to make ASR insensitive to noise it is crucial that these distributions are similar in the training and testing condition. Mostly, it is attempted to compensate for the impact of noise by estimating the noise characteristics from the signal. In this paper we explore the feasibility of a new method to increase noise robustness: We try to exploit a priori knowledge stored in clean speech models. Using Mel bank log-energy features, recognition is done by ignoring the model components for features that contained little energy during training. This strategy aims at recognition results that are determined more strongly by the match in the high-energy rather than by the mismatch in the low-energy model components. Application of the new method to clean speech data confirms that discarding components below a certain energy threshold does not deteriorate recognition performance. Experiments with noisy data, however, show that performance gains are relatively small. This paper explains why that is the case and why, despite the limited success, the outcomes suggest that the method still could prove to be a valuable addition to data-driven methods like (bounded) marginalisation