Recently, I did some comparisons among the ctc_greedy_search, ctc_prefix_beam_search, attention and attention_rescore.

1. Test machine:
   GPU: NVIDIA GeForce RTX 4090
   CPU Intel(R) Xeon(R) Platinum 8336C @ 2.30GHz.
2. Test config
   Beam size: 10
   model: u2++_conformer_exp_aishell
   test set: aishell(7176 utts, ~5hours).
   platform: python.
   batchsize: 1
   chunksize: full
3. Results:

So I have some questions:

1. Firstly, are the results reasonable?
   the speed is ctc_greedy_search > attention > ctc_prefix_beam_search > attention_rescore.

2. why the attention is faster than ctc_prefix_beam_search?
   it is because the u2++_conformer_exp_aishell decoder is simple, which causes the attention is faster than ctc_prefix_beam_search?

3. it seems attention_resore needs to do the ctc_prefix_beam_search first, which will slow the speed.

Thanks!