Latent Sequence Decompositions

Citations Over Time

Abstract

We present the Latent Sequence Decompositions (LSD) framework. LSD decomposes sequences with variable lengthed output units as a function of both the input sequence and the output sequence. We present a training algorithm which samples valid extensions and an approximate decoding algorithm. We experiment with the Wall Street Journal speech recognition task. Our LSD model achieves 12.9% WER compared to a character baseline of 14.8% WER. When combined with a convolutional network on the encoder, we achieve 9.6% WER.

Related Papers

• → Long Short-Term Memory(1997)95,038 cited
• → Connectionist temporal classification(2006)5,326 cited
• → Listen, attend and spell: A neural network for large vocabulary conversational speech recognition(2016)2,316 cited