Is it possible to bring back old Kaldi's chain-model style training back to Icefall

[davidlin409]

Recently I need to port Kaldi's chain model training to Icefall.

Alhough Icefall's method is with better result, but we are having very not-normal setup, such that Icefall's training method (based on CTC) does not work for us. However, Icefall have lots of new algorithm, which is very hard to port one-by-one to Kaldi.

Therefore, I am trying to use alignment file from Kaldi, to feed those alignment-data into Icefall training flow, and use chain-style training method; which is, to convert alignment (originally it is lattice in Kaldi) to training-FST-graph, and use it inside Icefall training.

However, converting alignment from Kaldi to Icefall, while putting alignment into Icefall-training-FST-graph, is very confusing. For example, I am not sure how I can turn Kaldi alignment into Icefall-training-FST graph, where original MMI-graph need to add self-eps-loop to P-gram.

Is there any documentation, or any example to port Kaldi's chain model training method back to Icefall?

PS: Chain model here is the non-E2E version of chain model in old Kaldi, not the E2E one.

[yfyeung]

Currently, there is no method.

[danpovey]

I think maybe because you are using a too-old recipe? Can you try the recipe in librispeech/ASR/zipformer?
If you have enough data, e.g. a few hundred hours at least, I think it should be better than chain model.