Method for adaptive on-line data fusion in multi-channel automatic speech recognition systems

In this paper describes a method for adaptive, parameter-independent, on-line channels' weight estimation, based on entropy in the output probabilities from ANN classifiers, rather than the noise level or SNR estimation. A recursive formula for channel combination calculation, based on the approximation of all possible channels combination, is deduced. The proposed method has been used to develop a multi-channel distributed speech recognition (DSR) system. From experiments a conclusion can be drawn that the use of the proposed method results in an absolute system accuracy improvement of 12.4% in comparison with the base one-channel DSR system from the ETSI Aurora Project.