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.
@article{arxiv.1610.03035,
title = {Latent Sequence Decompositions},
author = {William Chan and Yu Zhang and Quoc Le and Navdeep Jaitly},
journal= {arXiv preprint arXiv:1610.03035},
year = {2017}
}