Thanks for creating a repo for multitask learning in Kaldi. It is much needed.
I have two questions regarding MTL in Kaldi.
I have speech datasets from two different domains and I want to add another branch in this TDNN architecture which would classify prefinal-l into those domains. I think I will have to modify the egs to contain the domain labels, but how will this label be provided to this branch during training?
I want to combine ASR training with speaker classification (x-vector) architecture. I think the above egs modification will suffice here too.
Could you please provide me some guidelines how to go about these?
Dear @JRMeyer ,
Thanks for creating a repo for multitask learning in Kaldi. It is much needed. I have two questions regarding MTL in Kaldi.
prefinal-l
into those domains. I think I will have to modify the egs to contain the domain labels, but how will this label be provided to this branch during training?Could you please provide me some guidelines how to go about these?
Thanks! Brij