In this paper, we report state-of-the-art results on LibriSpeech among end-to-end speech recognition models without any external training data. Our model, Jasper, uses only 1D convolutions, batch normalization, ReLU, dropout, and residual connections. To improve training, we further introduce a new layer-wise optimizer called NovoGrad. Through experiments, we demonstrate that the proposed deep architecture performs as well or better than more complex choices. Our deepest Jasper variant uses 54 convolutional layers. With this architecture, we achieve 2.95% WER using a beam-search decoder with an external neural language model and 3.86% WER with a greedy decoder on LibriSpeech test-clean. We also report competitive results on the Wall Street Journal and the Hub5'00 conversational evaluation datasets.
Cite
@article{arxiv.1904.03288,
title = {Jasper: An End-to-End Convolutional Neural Acoustic Model},
author = {Jason Li and Vitaly Lavrukhin and Boris Ginsburg and Ryan Leary and Oleksii Kuchaiev and Jonathan M. Cohen and Huyen Nguyen and Ravi Teja Gadde},
journal= {arXiv preprint arXiv:1904.03288},
year = {2019}
}