A resource-constrained HCRF modeling for a large-scale speaker identification task

This paper proposes an efficient algorithm on the training of hidden conditional random fields (HCRFs) for large-scale speaker recognition in which a speaker identification task with around 1000 speakers is investigated. HCRFs are a type of direct models in pattern recognition and thus iterative procedures are usually required to estimate the model parameters. The key method in this paper is to perform the optimization alternatively between the native direct models (HCRFs) and their approximated generative models (i.e., the equivalent HMMs of HCRFs). By this framework, a constrained optimization method is proposed which makes the training process of speaker models consumes less computation resources than the one by the conventional HCRF training scheme. The experimental results indicate that the proposed method enjoys potential benefits for development in resource-constrained speaker recognition.