Deep-hidden Conditional Neural Fields for Continuous Phoneme Speech Recognition

We have proposed Hidden Conditional Neural Fields (HCNF) for automatic speech recognition and shown the effectiveness by continuous phoneme recognition experiments on the TIMIT and the Japanese ASJ+JNAS corpora. In this paper, we propose to use an observation function with a deep structure in HCNF. The proposed deep observation function enables to use the deep neural networks in HCNF, which have recently achieved remarkable success in Automatic Speech Recognition. We call the HCNF with the observation function with a deep structure Deep-HCNF. Experimental results of continuous phoneme speech recognition on the TIMIT and the ASJ+JNAS corpora showed that the Deep-HCNFs with monophone structure outperformed traditional tied-state triphone HMMs trained in MPE manner.