Current browse context:
eess
Change to browse by:
References & Citations
Computer Science > Computation and Language
Title: A Novel Topology for End-to-end Temporal Classification and Segmentation with Recurrent Neural Network
(Submitted on 10 Dec 2019)
Abstract: Connectionist temporal classification (CTC) has matured as an alignment free to sequence transduction and shows competitive for end-to-end speech recognition. In the CTC topology, the blank symbol occupies more than half of the state trellis, which results the spike phenomenon of the non-blank symbols. For classification task, the spikes work quite well, but as to the segmentation task it does not provide boundaries information. In this paper, a novel topology is introduced to combine the temporal classification and segmentation ability in one framework.
Link back to: arXiv, form interface, contact.