Created by: Jun-Fan

In the original paper of DeepSpeech2, lookahead convolution is introduced for unidirectional model to keep accuracy, lower the latency compared to bidirectional model. And the best English models in the paper is 3 layers of unidirectional GRU with lookahead convolution. So, it seems that lookahead convolution is great for real-time deploy, but why lookahead convolution is not implemented in this repository?

Thanks!