[Ready to merge] Pruned Transducer Stateless2 for WenetSpeech (char-based)

[luomingshuang]

Actually, there is an existing PR #314 with pruned transducer stateless2 for wenetspeech. But that PR can be regarded as a draft. This PR aims to merge into the master. Of course, the process trained with the L subset is still going on. (The results are better than other methods at present. And the performance still gets a little improvement. I will update the results on RESULT.md) It doesn't influence the code review for this PR. The greedy_search_new and modified_beam_search_new are from the PR #358.
Results:

Subset_for_training	decoding_method	epoch	avg	pruned RNN-T (test-net/test-meet)	reworked model (dev/test-net/test-meet)
M	greedy_search	29	11	-/-	10.40/11.31/19.64
M	modified_beam_search	29	11	-/-	9.85/11.04/18.20
M	fast_beam_search	29	11	-/-	10.18/11.10/19.32
M	/	/	/	kaldi	9.81/14.19/28.22
S	greedy_search	29	24	-/-	19.92/25.20/35.35
S	modified_beam_search	29	24	-/-	18.62/23.88/33.80
S	fast_beam_search	29	24	-/-	19.31/24.41/34.87
S	/	/	/	kaldi	11.70/17.47/37.27
L	greedy_search	10	2	-/-	7.80/8.78/13.49
L	greedy_search_new	10	2	-/-	7.80/8.76/13.50
L	modified_beam_search	10	2	-/-	7.76/8.82/13.41
L	modified_beam_search_new	10	2	-/-	7.76/8.72/13.41
L	fast_beam_search	10	2	-/-	7.94/8.73/13.81
L	/	/	/	kaldi	9.07/12.83/24.72
L	/	/	/	wenet	8.88/9.70/15.59
L	/	/	/	espnet	9.70/8.90/15.90

[luomingshuang]

The best results trained with L subset (better than other public results at present):

	dev	test-net	test-meeting	comment
greedy search	7.80	8.75	13.49	--epoch 10, --avg 2, --max-duration 100
modified beam search (beam size 4)	7.76	8.71	13.41	--epoch 10, --avg 2, --max-duration 100
fast beam search (set as default)	7.94	8.74	13.80	--epoch 10, --avg 2, --max-duration 1500

[csukuangfj]

Please resolve the conflicts.

[luomingshuang]

Done.

[csukuangfj]

Could you also add the pretrained models and decoding results for the M and S subset to huggingface?

Also, could you upload the training logs to huggingface?

[luomingshuang]

OK, no problem.

[csukuangfj]

Thanks!

[pingfengluo]

@luomingshuang how long it takes to complete wenetspeech training with L set?

[luomingshuang]

You can find the tensorboard log here. And according to this log, it takes about 21 days 16 hours and a half for 17 epochs. According to my results, I think it can be stopped when training for 11 epochs.

[luomingshuang]

https://tensorboard.dev/experiment/wM4ZUNtASRavJx79EOYYcg/#scalars

[pingfengluo]

tks