Speed of Light Exact Greedy Decoding for RNN-T Speech Recognition Models on GPU