ASR model and HCLG.fst LibriSpeech

on


Many final models final.mdl in various folders, which one is the final one we need?


During training the LibriSpeech I noticed that we are getting many final models final.mdl in various folders

ubuntu@ip-172-31-6-144:~/kaldi/egs/librispeech/s5/exp$ ls -alh
total 136K
drwxrwxr-x 22 ubuntu ubuntu 4.0K Dec 31 02:29 .
drwxrwxr-x  7 ubuntu ubuntu 4.0K Jan  8 02:38 ..
drwxrwxr-x  5 ubuntu ubuntu 4.0K Dec 31 19:06 chain_cleaned
drwxrwxr-x  9 ubuntu ubuntu 4.0K Dec 28 17:01 make_mfcc
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:01 mono
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:01 mono_ali_5k
drwxrwxr-x 10 ubuntu ubuntu 4.0K Dec 31 02:29 nnet3_cleaned
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:05 tri1
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:06 tri1_ali_10k
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:14 tri2b
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:15 tri2b_ali_10k
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:25 tri3b
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:32 tri3b_ali_clean_100
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 11:58 tri4b
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 13:41 tri4b_ali_clean_460
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 15:55 tri5b
drwxrwxr-x  3 ubuntu ubuntu 4.0K Dec 28 18:44 tri5b_ali_960
drwxrwxr-x 24 ubuntu ubuntu 4.0K Dec 29 01:43 tri6b
drwxrwxr-x  3 ubuntu ubuntu  20K Dec 30 04:29 tri6b_ali_cleaned
drwxrwxr-x 24 ubuntu ubuntu  20K Dec 30 13:20 tri6b_cleaned
drwxrwxr-x  3 ubuntu ubuntu  20K Dec 30 21:05 tri6b_cleaned_ali_train_960_cleaned_sp
drwxrwxr-x  6 ubuntu ubuntu 4.0K Dec 30 02:55 tri6b_cleaned_work

Most of those will just be GMM models, they are used in the system building process. The one in exp/chain_cleaned/ is the one that counts

Where to find the final ASR model ?

Look at /exp/chain_cleaned/tdnn_1d_sp folder for LibriSpeech final model

How do we get H for HCLG.fst ?

The information from it should come from data/lang_chain/topo or data/lang_chain_sp/topo

it is not actually represented as an FST.

How do we get C for HCLG.fst ?

Also C is not represented as a physical FST,

there is a program fstcomposecontext that composes with it programatically

How do we get L for HCLG.fst ?

L.fst can be found in data/lang or data/lang_nosp

How do we get G for HCLG.fst ?

G.carpa can be found in data/lang_test_tglarge or data/lang_nosp_test_fglarge













































Comments

Post a Comment