I cant find native_client.tar.xz [DeepSpeech 0.6.1]

Archives DeepSpeech
#1

Hello,

I am following the tutorial french model trainning but I am stucked in the part of create the trie file.
I was looking for the native_client.tar.xz where it is the binary of the generate trie, but I just can’t find it. My native_client folder has only the generate_trie.cpp file but no tar.xz.
I also tried to build the binary with the command:

bazel build --workspace_status_command="bash native_client/bazel_workspace_status_cmd.sh" --config=monolithic -c opt --copt=-O3 --copt="-D_GLIBCXX_USE_CXX11_ABI=0" --copt=-fvisibility=hidden //native_client:libdeepspeech.so //native_client:generate_trie

Extracting Bazel installation...
ERROR: The 'build' command is only supported from within a workspace (below a directory having a WORKSPACE file).
See documentation at https://docs.bazel.build/versions/master/build-ref.html#workspace

But the above error appears. (executing in the native_client folder)

Can someone help me?

Python: 3.7.3
SO: Debian buster
GPU: RTX 2060 SUPER
Using common voice Portugues dataset.

For training I am using the commando below:

./DeepSpeech.py --checkpoint_dir /home/deep_train/checkpoint --train_files /home/deep_train/portugues/clips/train.csv --dev_files /home/deep_train/portugues/clips/dev.csv --test_files /home/deep_train/portugues/clips/test.csv --export_dir /home/deep_train/portugues --alphabet_config_path /home/deep_train/portugues/alphabet.txt --lm_binary_path /home/deep_train/portugues/clips/lm.binary --train_batch_size 8 --test_batch_size 8 --dev_batch_size 8 --export_batch_size 1 --epochs 75 > /home/train.log &

And when it is finished and I try to transcribe any audio, I only get “a”, “e” and “o” letters. Command to transcribe, for example:
deepspeech --model /home/deep_train/portugues/output_graph.pb --audio /home/deep_train/portugues/clips/common_voice_pt_19887570.wav

I appreciate any help with the trie file.
Best regards.

#2

Have you tried looking at the release page? https://github.com/mozilla/DeepSpeech/releases/tag/v0.6.1

1 Like
#3

Have you properly followed the documentation to build it ?

#4

This is not actionable if you don’t share your training log. How can we know what your model learnt ? What was the test set result ?

You might want to try and join efforts on building on top of this Docker https://github.com/Common-Voice/commonvoice-fr/blob/master/DeepSpeech/Dockerfile.train like Italian community does: https://github.com/MozillaItalia/DeepSpeech-Italian-Model

#5

What material did you use to build your LM ? Also, you pass no trie, but I guess it is because of generate_trie.

1 Like
#6

I builded the kenlm from source.
I am using the Common Voice portuguese dataset, for lm.binary I used:

cat dev.csv >> vocabulary.txt && cat test.csv >> vocabulary.txt && cat train.csv >> vocabulary.txt
/home/deep_train/kenlm/build/bin/lmplz --text vocabulary.txt --arpa words.arpa --o 3
/home/deep_train/kenlm/build/bin/build_binary -T -s words.arpa lm.binary

cat is for create a big file with all the dev, train and test csv.

#7

I guess it might work with some pre-filtering. But here you are also passing CSV metadata, filenames and filesize. That’s likely going to confuse a lot your LM.

For example, this should be a bit better.

$ tail -n -1 dev.csv | cut -d ',' -f 3 > vocabulary.txt
#8

I get it, its only the text in this file, not the wav filename or file size.
Thanks.

#9

Hello again, I will post here the log from the test part. The command to train the model:

./DeepSpeech.py --checkpoint_dir /home/deep_train/checkpoint --train_files /home/deep_train/portugues/clips/train.csv --dev_files /home/deep_train/portugues/clips/dev.csv --test_files /home/deep_train/portugues/clips/test.csv --export_dir /home/deep_train/portugues --alphabet_config_path /home/deep_train/portugues/alphabet.txt --automatic_mixed_precision=True --early_stop=False --lm_binary_path /home/deep_train/portugues/clips/lm.binary --lm_trie_path /home/deep_train/portugues/clips/trie --train_batch_size 32 --test_batch_size 16 --dev_batch_size 16 --export_batch_size 1 --learning_rate 0.00095 --epochs 100 > /home/train.log &

I used early stop false because I wanted to do all the iterations. But the log result are still very bad:

Test on /home/deep_train/portugues/clips/test.csv - WER: 0.866583, CER: 0.601828, loss: 107.132584

WER: 2.500000, CER: 0.866667, loss: 34.759850

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19435022.wav
  • src: “ativar legendas”
  • res: “o de vale de no”

WER: 2.333333, CER: 0.629630, loss: 62.630676

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19401141.wav
  • src: “pode solicitar participação”
  • res: “o sol e a aros e as”

WER: 2.000000, CER: 0.636364, loss: 21.256580

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19312911.wav
  • src: “notarização”
  • res: “duraria a”

WER: 2.000000, CER: 0.700000, loss: 27.122267

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_20059720.wav
  • src: “brevemente”
  • res: “de onde”

WER: 2.000000, CER: 0.750000, loss: 30.222834

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19403802.wav
  • src: “é lógico”
  • res: “e o de o”

WER: 2.000000, CER: 0.642857, loss: 33.146614

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_20017611.wav
  • src: “muito dinheiro”
  • res: “o de de mero”

WER: 2.000000, CER: 0.611111, loss: 33.159065

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19375886.wav
  • src: “eu chamei benjamin”
  • res: “eu te a me de a vi”

WER: 2.000000, CER: 0.750000, loss: 36.686123

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19458070.wav
  • src: “outra plataforma”
  • res: “de prata sua mar”

WER: 2.000000, CER: 0.750000, loss: 49.787704

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19338110.wav
  • src: “quanto foi novamente”
  • res: “o sino na a me o”

WER: 2.000000, CER: 0.695652, loss: 51.571831

  • wav: file:///home/deep_train/portugues/clips/common_voice_pt_19420984.wav
  • src: “desporto associativismo”
  • res: “são um ano teve”

I Exporting the model…
I Models exported at /home/deep_train/portugues

The common voice Portuguese dataset has only 27h validated. is it so little that he can’t even learn?
Also, I used 100 iterations, but I am not sure it can improve so much(since 75 iter is recommended)
I fixed the trie and he is clean now, only the phrases.

#10

Oh yeah, with 27h you can’t expect anything really. Your best bet here would be to use transfer learning. This is on master, though.