SURT recipe for AMI and ICSI

[desh2608]

This recipe has a dependency on #1126.

It trains a SURT model jointly on AMI and ICSI. We get the following results on utterance group based evaluation. SURT-base contains 26.7M params and SURT-large contains 37.9M params.

AMI

Model	IHM-Mix	SDM	MDM
SURT-base	39.8	65.4	46.6
+ adapt	37.4	46.9	43.7
SURT-large	36.8	62.5	44.4
+ adapt	35.1	44.6	41.4

ICSI

Model	IHM-Mix	SDM
SURT-base	28.3	60.0
+ adapt	26.3	33.9
SURT-large	27.8	59.7
+ adapt	24.4	32.3

• Pre-trained models: https://huggingface.co/desh2608/icefall-surt-ami-dprnn-zipformer

• Training logs:

  • surt_base: https://tensorboard.dev/experiment/8awy98VZSWegLmH4l2JWSA/
  • surt_base_adapt: https://tensorboard.dev/experiment/aGVgXVzYRDKbGUbPekcNjg/
  • surt_large: https://tensorboard.dev/experiment/ZXMkez0VSYKbPLqRk4clOQ/
  • surt_large_adapt: https://tensorboard.dev/experiment/WLKL1e7bTVyEjSonYSNYwg/

[desh2608]

@yaozengwei Thanks for merging the LibriCSS recipe. This recipe follows the same idea, but it is for AMI and ICSI. Most files are symlinks of the ones in LibriCSS.

[yaozengwei]

@desh2608 Is it ready to merge?

[desh2608]

@desh2608 Is it ready to merge?

Yes!