issues/doubts about the current L.fst

[csukuangfj]

Some issues and doubts about the current method to generate L.fst:

• ~Since there is no G, do we still need the disambiguation symbols?~

• What is the purpose of adding optional silence at the end of each word? Only OOV words are mapped to the silence phone SIL and I am afraid there is not enough training data for SIL.

• ~There is no blank symbol at all. Why is it not used? It is common to add a blank symbol to the topology in the literature, I think.~

• No test code.

[qindazhu]

RE 1 and 3.

• We use L.fst for training, there's no diambig symbols in it.
• we add self-loops (blank) in training and decoding.

[csukuangfj]

And another issue:

• Any reason to NOT change the topology to support repeated symbols?

[danpovey]

• For the optional silence, we'll have to see if it's helpful in any scenario. I think once we start using LF_MMI as part of the objective function it will become useful.
• RE changing the topology to support repeated symbols, there isn't a particular reason, you could go ahead and try it.

[csukuangfj]

thanks, will try it.

[pzelasko]

I am pretty sure that the disambiguation symbols in the phone symbol table are also present in the acoustic model output layer, as my changes didn't address that yet. I am not sure if them being present in the phone symbol table means that they are also present in L.

[danpovey]

In graph.py this line LG.labels[LG.labels >= labels_disambig_id_start] = 0 takes care of it.