We propose a fully convolutional sequence-to-sequence encoder architecturewith a simple and efficient decoder. Our model improves WER on LibriSpeechwhile being an order of magnitude more efficient than a strong RNNbaseline. Key to our approach is a time-depth separable convolutionblock which dramatically reduces the number of parameters in the modelwhile keeping the receptive field large. We also give a stable andefficient beam search inference procedure which allows us to effectivelyintegrate a language model. Coupled with a convolutional language model,our time-depth separable convolution architecture improves by morethan 22% relative WER over the best previously reported sequence-to-sequenceresults on the noisy LibriSpeech test set.
@inproceedings{hannun19_interspeech, title = {Sequence-to-Sequence Speech Recognition with Time-Depth Separable Convolutions}, author = {Awni Hannun and Ann Lee and Qiantong Xu and Ronan Collobert}, year = {2019}, booktitle = {Interspeech 2019}, pages = {3785--3789}, doi = {10.21437/Interspeech.2019-2460}, issn = {2958-1796},}Cite as:Hannun, A., Lee, A., Xu, Q., Collobert, R. (2019) Sequence-to-Sequence Speech Recognition with Time-Depth Separable Convolutions. Proc. Interspeech 2019, 3785-3789, doi: 10.21437/Interspeech.2019-2460