Besides exploiting the shortcomings of the HMMs, another motivation to use such an approach is the difference in the recognition rates when considering only the best word hypothesis and the first 10-best word hypotheses, which is greater than 15 percent for an 80,000-word lexicon. This differenceis attributed to thepresence of very similarwordsinฃ the lexicon that may differ only by one or two characters. The proposed verification approach is expected to better discriminate between characters and to alleviate this problem.