the output of LSTM

[wqn628]

First, thanks for your help all the time. And I have been being confused by the modeled units all the time .For instance : The unit.txt And I wonder why we should model the first phone and the second phone ,Actually,both of them don't exist in my training label.Can I delete them and not model them ? any help would be appreciated.

[wqn628]

@yajiemiao

[wqn628]

why should we add noise to the lexicon(noises phonemes to the units) ?

[yajiemiao]

if they truly don't exist in your training data, you can safely delete them but caution that by default, Eesen maps OOV words in your training transcripts to

[wqn628]

Thanks a lot. In addition, the Essen makes model for mono-phone directly, can tri-phones be the model units in essen ? @yajiemiao . As the previous acoustic model(GMM_HMM DNN/LSTM_HMM),the tri-phone have outperformed a lot than mono-phone.

[chenzhehuai]

using tri-phone as the model unit in essen is possible, u might further generate context label (fstcomposecontext) as in HMM system, and replace tri-phone label in T.fst. The final WFST changes into T\circ C\circ LG

[wqn628]

sorry ,i don't got it. you mean that I should generate the tri-phone by the hybird pipeline or by the commmand ----"fstcomposecontext".the first or the second ?@chenzhehuai

[chenzhehuai]

clustered tri-phone should be generated from hybrid system through clustering; while context in WFST can be generated by fstcomposecontext with extra mapping from tri-phone to clustered tri-phone

[yajiemiao]

An even simpler way is to generate forced alignment with the GMM-HMM, and take the CD states as CI CTC labels. With this, there is no need to consider context dependency in decoding. I didn't do such an experiment, so not sure how this could work in practice.

[wqn628]

hello,in the stage of decoding, the problem occur as follows:

can you tell what had happened and how I can solve them? thanks a lot. @yajiemiao