Applying conditional random fields on Chinese syllable recognition

Hidden Markov model (HMM) is successfully used in speech recognition. However, there is an unavoidable flaw in assuming strong independence for sequences labeling in HMM. While conditional random fields (CRFs) can relax this assumption for HMM, and can also solve the label bias problem efficiently. In this paper, we investigate CRFs for Chinese syllable recognition in continuous speech due to its advantages. The experiments show that the syllable label CRF is able to achieve performance comparable to phone-based HMM.