Getting wrong output from the trained model

DeepSpeech
#1

Hi,

I am able to train a model on language(Hindi) but getting wrong output from the model when passing a sample ‘wav’ file. Have tried many times but no luck.

Below is the command I am using to train the mode:-

python -u DeepSpeech.py --train_files data/flac_csv2.csv --dev_files data/flac_csv2.csv --test_files data/flac_csv2.csv --train_batch_size 10 --dev_batch_size 10 --test_batch_size 10 --n_hidden 1028 --export_dir model_export --epoch 75 --checkpoint_dir /home/sp/.local/share/deepspeech/flac_new2 > flac_new1.log

We are providing the 4600 input wav files and their transcriptions in the csv file. Model is getting trained successfully but returning wrong output.

Can you please guide me through the steps to train the model as I might be missing any step.

Also please do let me know,if we need to create our own ‘lm.library’ file for specific language.

Any help would be appreciated.

Thanks in advance… :slight_smile:

#2

We would need more context. 4600 wav might not be enough data, you changed n_hidden to 1028, that might also not be good, you might be doing too many epochs. Without more context and the training log, we can’t help. Is the wrong data not learning enough? Overfitting? Can’t tell.

#3

Thanks lissyx for the reply. :slight_smile:

PFB output of the logs:

Preprocessing [‘data/flac_csv.csv’]
Preprocessing done
Preprocessing [‘data/flac_csv.csv’]
Preprocessing done
W Parameter --validation_step needs to be >0 for early stopping to work
I STARTING Optimization
I Training epoch 0…
I Training of Epoch 0 - loss: 289.471059
I Training epoch 1…
I Training of Epoch 1 - loss: 303.551886
I Training epoch 2…
I Training of Epoch 2 - loss: 303.208348
I Training epoch 3…
I Training of Epoch 3 - loss: 303.340517
I Training epoch 4…
I Training of Epoch 4 - loss: 303.261908
I Training epoch 5…
I Training of Epoch 5 - loss: 303.110338
I Training epoch 6…
I Training of Epoch 6 - loss: 303.026394
I Training epoch 7…
I Training of Epoch 7 - loss: 303.048171
I Training epoch 8…
I Training of Epoch 8 - loss: 302.989858
I Training epoch 9…
I Training of Epoch 9 - loss: 303.014609
I Training epoch 10…
I Training of Epoch 10 - loss: 302.963002
I Training epoch 11…
I Training of Epoch 11 - loss: 302.966895
I Training epoch 12…
I Training of Epoch 12 - loss: 302.946233
I Training epoch 13…
I Training of Epoch 13 - loss: 302.952651
I Training epoch 14…
I Training of Epoch 14 - loss: 302.931431
I FINISHED Optimization - training time: 0:59:37
Preprocessing [‘data/flac_csv.csv’]
Preprocessing done
Computing acoustic model predictions…
Decoding predictions…
Test - WER: 0.970377, CER: 98.243197, loss: 302.404297

WER: 1.000000, CER: 6.000000, loss: 27.665541

  • src: “a story”
  • res: "i "

WER: 1.000000, CER: 6.000000, loss: 27.665541

  • src: “a story”
  • res: "i "

WER: 1.000000, CER: 6.000000, loss: 31.257378

  • src: “oh emil”
  • res: "i "

WER: 1.000000, CER: 4.000000, loss: 32.887798

  • src: “ay me”
  • res: "i "

WER: 1.000000, CER: 4.000000, loss: 32.887798

  • src: “ay me”
  • res: "i "

WER: 1.000000, CER: 8.000000, loss: 33.150005

  • src: “direction”
  • res: "i "

WER: 1.000000, CER: 5.000000, loss: 34.121330

  • src: “venice”
  • res: "i "

WER: 1.000000, CER: 8.000000, loss: 34.133373

  • src: “verse two”
  • res: "i "

WER: 1.000000, CER: 7.000000, loss: 34.345585

  • src: “indeed ah”
  • res: "i "

WER: 1.000000, CER: 7.000000, loss: 34.345585

  • src: “indeed ah”
  • res: "i "

I Exporting the model…
I Models exported at model_export

The output is same in all the cases…

Its been almost 5 days I am stuck in this thing as I am daily trying a different approach to train the model with the help of the related issues and the replies on them but no luck.
Can you please help me with the steps to train the model in my language. I might be missing something which i am not able to trace down. :frowning:

Thanks…

#4

I had the same output, my data was in 8khz format, I thought maybe this is main reason.

Hope we can solve this problem

Here the is the my output:

Computing acoustic model predictions…
100% (285 of 285) |######################| Elapsed Time: 0:08:37 Time: 0:08:37
Decoding predictions…
100% (285 of 285) |######################| Elapsed Time: 0:07:52 Time: 0:07:52
Test - WER: 0.993192, CER: 66.797807, loss: 326.212860

WER: 1.500000, CER: 17.000000, loss: 172.890213

  • src: “оформили подскажите”
  • res: “в т п”

WER: 1.500000, CER: 14.000000, loss: 197.776184

  • src: “қанша переводите”
  • res: “в т п”

WER: 1.333333, CER: 16.000000, loss: 208.935623

  • src: “это родственники ваш”
  • res: “в т е е”

WER: 1.000000, CER: 20.000000, loss: 1.283422

  • src: “номер телефона подскажите”
  • res: “е т и”

WER: 1.000000, CER: 7.000000, loss: 4.558431

  • src: “да да да”
  • res: “а”

WER: 1.000000, CER: 15.000000, loss: 4.634631

  • src: “да ночи ежедневно”
  • res: “не”

WER: 1.000000, CER: 10.000000, loss: 7.905017

  • src: “да это тальго”
  • res: “эль”

WER: 1.000000, CER: 19.000000, loss: 7.950384

  • src: “на седьмое чего декабря”
  • res: “в т ер”

WER: 1.000000, CER: 18.000000, loss: 9.858935

  • src: “ну почему да причина”
  • res: "ни "

WER: 1.000000, CER: 21.000000, loss: 10.437287

  • src: “на шестнадцатое декабря”
  • res: “не”

I Exporting the model…
I Models exported at /home/dulan/models

#5

Are you still having this problem ?