search

jzlianglu / pykaldi2

Yet another speech toolkit based on Kaldi and PyTorch
MIT License
173 stars 33 forks source link

Problems about HCLG.fst needed by "train_transformer_se.py" #20

Closed SiyuanWei closed 4 years ago

SiyuanWei commented 4 years ago

Hi, I'm implementing your paper "A TRANSFORMER WITH INTERLEAVED SELF-ATTENTION AND CONVOLUTION FOR HYBRID ACOUSTIC MODELS", and I'v got some questions bothering me. In kaldi setup, sequence training need to create a phone-level language model and denominator fst, which is called HCP in this blog (https://desh2608.github.io/2019-05-21-chain/). In your code, I find that the script "train_transformer_se.py" needs a direction that contains HCLG.fst. Is the HCLG.fst needed here is equal to the HCP builded based on a phone-level LM.?

jzlianglu commented 4 years ago

Hi, you are talking about LF-MMI training of Kaldi. The SE training script for transformer model is using lattice-based approach. For LF-MMI training, please refer to train_chain.py in the "bin" folder. Please note that, the chain model training script only uses batch size as 1. I will update the recipe to support large batch size in the near future.

SiyuanWei commented 4 years ago

Oh, Thanks !!!