Connect word ends to initial state for context-independent allophones

[SimBe195]

For context-independent allophones (i.e. history_length = future_length = 0) the word-end states were not connected to the initial state in the ClassicTransducerBuilder which results in a broken automaton and a crash during training with fast_bw loss (e. g. CTC).
A fix for this behaviour had already been implemented in the original label_sync_decoding branch of @ZhouW321 but was missed when porting to github RASR. This PR re-applies the original fix.

[mmz33]

Hey @curufinwe, we need this urgently for some experiments so could you please review this soon if possible? thank you.

[curufinwe]

I'm currently not super familiar with this part of the code. Could you attach plots of a simple automaton before and after your fix? Thanks!

[Marvin84]

Yes, me and @michelwi are aware of this problem, since the FSA that uses allophones without right and left history is broken. I agree that there should be more documentation and test that confirms the correctness of the resulting FSA and I also suggest that this PR includes SVGs of an example.

I personally find that this solution is not mathematically incorrect. The classic FST creation is constructing an HCLG graph, and the developer of seq2seq framework needs an HLG graph, but by making all right and left context to be equal to # you are not creating an HLG graph. The correct solution should have been skipping the C composition.

[SimBe195]

I'm currently not super familiar with this part of the code. Could you attach plots of a simple automaton before and after your fix? Thanks!

@curufinwe Without the fix the words in a segment are not connected which causes the allophoneStateToPhonemeTransducer to not have a final state, so after trimming the automaton will be empty (and further processing steps crash).
For visualization I have selected a short segment containing only the words "THIS WEEK" and I have removed the Fsa::trim so that you can see the automaton properly.

This is the allophoneStateToPhonemeTransducer without the fix:

And this is the allophoneStateToPhonemeTransducer with the fix:

[albertz]

Note that this fix here is needed to fix #68 and rwth-i6/returnn#1456.

[Marvin84]

Regarding the allophoneStateToPhonemeTransducer I am attaching here the resulting FSA from our standard HMM with the proper composition. As you can see, there are no dead-end nodes, meaning all paths will finish in one final node. I believe this is one first trivial condition for a proper FSA definition. The question is how this current composition affects the optimization and convergence when for example doing full-sum training.