conformer-ctc
A Appendix
We evaluate the final models on both clean and other datasets using CTC greedy decoding. Here, we use CUDA 11.5 and Tensorflow 2.5, and test with the largest In both cases, we used the same Sentence-Piece tokenizer as Librispeech training. All the other training recipes are the same as A.1. As can be seen in the table, the general trend aligns with the LibriSpeech results in Tab.