Sequential routing framework: Fully capsule network-based speech recognition

Capsule networks (CapsNets) have recently gotten attention as a novel neural architecture. This paper presents the sequential routing framework which we believe is first method to adapt CapsNet-only structure sequence-to-sequence recognition. Input sequences are capsulized then sliced by window size. Each slice classified label at corresponding time through iterative mechanisms. Afterwards, losses computed connectionist temporal classification (CTC). During routing, required number of parameters can be controlled size regardless length sharing learnable weights across slices. We additionally propose dynamic algorithm replace traditional routing. The proposed technique minimize decoding speed degradation caused iterations since it operate in non-iterative manner without dropping accuracy. achieves 1.1% lower word error rate 16.9% on Wall Street Journal corpus compared bidirectional long short-term memory-based CTC networks. On TIMIT corpus, attains 0.7% phone 17.5% convolutional network-based (Zhang et al., 2016).