Many research found that Convolutional neural networks can reach near the state of art of Recurrent/LSTM/GRU neural networks ON many NLP/Audio process tasks, but with several times training speedup and less memory usage, then why we still use RNN? RNN should have more power on NLP/Audio, but too slow, maybe it has many incredible power we can't explore now, maybe it can reach near human intelligent in the future with more powerful computer power, in that future RNN maybe rediscover and revival like neural network now by DNN, but now can we just use CNN for nearly every thing in real product, no slow RNN?[1] An Empirical Evaluation of Generic CNN and RNN for Sequence Modeling:
http://arxiv.org/abs/1803.01271v2
[2] TextCNN VS RNN:
https://github.com/brightmart/text_classification#performance