Feature normalization based on non-extensive statistics for speech recognition