Phoneme Candidate Re-entry Modeling using Recognition Error Characteristics over Multiple HMM States