Size of the model after training

I tried the train as Shmyrev said.
It was difficult to set up the environment of kaldi but it was meaningful to me.

I Used this script
https://github.com/alphacep/vosk-api/blob/master/android/lib/build-vosk.sh#L112

When the learning is over, the result binary is larger than the vosk website.

Training results are as follows.
/exp/chain/tdnn/graph
HCLG.fst size is 151,552KB
phones.txt 2KB
words.txt 2,856KB

If the script you shared is correct, I think a similar size should come out.
I'm curious about your opinion.

Thanks, Moon.