Low resource end-to-end spoken language understanding with capsule networks

Designing a Spoken Language Understanding (SLU) system for command-and-control applications is challenging. Both Automatic Speech Recognition and Natural are language application dependent to great extent. Even with lot of design effort, users often still have know what say the it do they want. We propose use an end-to-end SLU that maps speech directly semantics can be trained by user through demonstrations. The teach new command uttering subsequently demonstrating its meaning alternative interface. will learn mapping from spoken task. dependency on also allows different languages non-standard or impaired as valid inputs. Teaching requires effort user, so crucial learns quickly. In this paper we capsule networks task, which believed data efficient. discuss two architectures using networks. analyse their performance compare them baseline systems, one based Non-negative Matrix Factorisation (NMF) has been successful task encoder-decoder approach. show in most cases network performs better than systems. Furthermore, demonstrate versatility architecture inferring speaker identity user’s word choice multitask learning.