Help on creating the model from the common voice mozilla

DeepSpeech
#1

Hi,

I was trying to create a new model from the common voice corpus (English).

As per the document I have downloaded the corpus en tar file.

Then extracted

bin/import_cv2.py --filter_alphabet path/to/some/alphabet.txt /path/to/extracted/language/archive

I am on this command -->

./DeepSpeech.py --train_files CommonVoice-2.0/corpus3-en/clips/train.csv --dev_files CommonVoice-2.0/corpus3-en/clips/dev.csv --test_files CommonVoice-2.0/corpus3-en/clips/test.csv --train_batch_size 12 --dev_batch_size 12 --test_batch_size 12 --learning_rate 0.0001 --epoch 95 --validation_step 5 --dropout_rate 0.30 --default_stddev 0.046875 --export_dir /opt/deepspeech/exportmodels --checkpoint /opt/deepspeech/checkpoint

But it is taking too long.

Please confirm once the command finishes it will bring like what inside the --> deepspeech-0.5.1-models

I assume it will create a new models file ( lm.binary, output_graph.pb, output_graph.pbmm, output_graph.tflite).

#2

Please share details, we do document that training is an intensive process, so without knowing your hardware as well as what you consider “too long”, we can’t do anything.

Don’t assume, read the documentation and the help.

#3

Hi @lissyx,

I am using Core i7 7th generation, 8gb ram, 4gb Nvidia graphics.

I will check the document and let you know if I have any doubts.

#4

Then no surprise, 250h of french audio takes ~4h for training on 2x RTX2080Ti GPUs.

1 Like
#5

@javi.rahman Just for the sake of completeness, can you make sure you are running tensorflow-gpu ? pip uninstall tensorflow && pip install --upgrade tensorflow-gpu==1.14.0 in your virtualenv.

1 Like