Explicit modelling of common acoustic features for character recognition

This paper presents a novel approach for robust, isolated character recognition. A major challenge of character recognition is that some characters are acoustically confusing and that no language modeling can be resorted to resolve the confusion. In the proposed approach, we attempt to explicitly model the common acoustic structures among different, confusing characters through state tying. As a result, decoding decision is made only by states modeling distinct sound segments. We first describe the training procedure of the new approach, then present recognition results from three character databases. Compared with the baseline system (which is a whole word/character model), the new approach is 45% better when evaluated using true telephone speech.