Add Blankskip to Zipformer+CTC

[yfyeung]

Accelerating RNN-T Training and Inference Using CTC guidance https://arxiv.org/pdf/2210.16481.pdf

with scaling Lconv

greedy_search

Model	test-clean	test-other	Decoding time(s) (3090)	Config
only decode	8.13	17.48	41.828	epoch 30 avg 13
both train and decode	5.74	15.46	39.717	epoch 30 avg 13
without blank skip	5.68	15.15	48.749	epoch 30 avg 13

modified_beam_search

Model	test-clean	test-other	Decoding time(s) (3090)	Config
only decode	7.88	17.01	100.744	epoch 30 avg 13
both train and decode	5.66	15.24	92.286	epoch 30 avg 13
without blank skip	5.64	14.93	219.133	epoch 30 avg 13

fast_beam_search

Model	test-clean	test-other	Decoding time(s) (3090)	Config
only decode	10.15	19.42	61.614	epoch 30 avg 13
both train and decode	7.72	17.51	65.103	epoch 30 avg 13
without blank skip	5.66	14.90	95.128	epoch 30 avg 13

without Lconv

greedy_search

Model	test-clean	test-other	Decoding time(s) (V100)	Config
only train	5.84	15.43	124.449	epoch 30 avg 13
both train and decode	5.92	15.59	112.946	epoch 30 avg 13
without blank skip	5.67	15.15	135.733	epoch 30 avg 13

modified_beam_search

Model	test-clean	test-other	Decoding time(s) (V100)	Config
only train	5.78	15.24	367.705	epoch 30 avg 13
both train and decode	5.86	15.31	153.612	epoch 30 avg 13
without blank skip	5.63	14.94	431.615	epoch 30 avg 13

CTC branch

Model	test-clean	test-other	Decoding time(s) (3090)	Config
Zipformer+CTC	6.28	16.76	56.284	epoch 30 avg 15
+ blank skip && scaling Lconv	6.24	16.97	54.181	epoch 30 avg 10