Skip to content

Latest commit

 

History

History

acoustic

README.md

English pre-trained model

This directory contain a trained english acoustic model : 67.600 steps of 30 files batches (3 mini-batchs of 10 files for each step).

Warnings :

  • This model use 3 layers, default is 5 in config file
  • Tensorflow's team love to change names of variables. This pre-trained model is "name compatible" with tensorflow 1.4 and maybe the next versions... If your result is incorrect you should set log_level to DEBUG and check if the TF variable names in the beginning of the log matches those listed in _add_saving_op in AcousticModel.py

Results on LibriSpeech's test-clean evaluation set (with max_input_seq_length : 3510):

  • CER : 15,2 %
  • WER : 42,4 %

Dataset

The dataset was built using Shtooka, LibriSpeech and TED-LIUM datasets

Training set contains :

  • LibriSpeech's train-clean-100
  • LibriSpeech's train-clean-360
  • LibriSpeech's train-other-500
  • Shtooka's eng-balm-emmanuel_flac.tar
  • Shtooka's eng-balm-judith_flac.tar
  • Shtooka's eng-balm-verbs_flac.tar
  • Shtooka's eng-wcp-us_flac.tar
  • Shtooka's eng-wims-mary_flac.tar
  • TED-LIUM's release 2

Test set contains :

  • LibriSpeech's test-clean

How to try it

First check that you have downloaded the lfs managed files. Execute this command in project's root dir :

$ git lfs pull

Then make sure to set those parameters in your config file :

[acoustic_network_params]
num_layers : 3
hidden_size : 1024
dropout_input_keep_prob : 0.8
dropout_output_keep_prob : 0.5
batch_size : 10
mini_batch_size : 3
learning_rate : 0.0003
lr_decay_factor : 0.33
grad_clip : 1
signal_processing : fbank
language : english
    
[general]
use_config_file_if_checkpoint_exists : True
steps_per_checkpoint : 100
steps_per_evaluation : 1000
checkpoint_dir : trained_models/english
    
[training]
max_input_seq_length : 3510
max_target_seq_length : 600
batch_normalization : False
dataset_size_ordering : False

Run the model on a "less than 15 seconds long" wav file of your choice

$ python3 stt.py --file data/LibriSpeech/dev-clean/2086/149220/2086-149220-0007.flac

On this example file from Librispeech dev set that the model never trained on you will obtain :

it now contained only shanetoclare his two wides and a solitary chicken

the original text being :

it now contained only chanticleer his two wives and a solitary chicken

Reproduce the learning phase

Put the training data directories in a "train" directory and set it in config file

training_dataset_dirs : data/Shtooka/train, data/LibriSpeech/train, data/TEDLIUM_release2/train

Put the test data in another directory and set it in config file

test_dataset_dirs : data/LibriSpeech/test

Launch training and wait...

$ python3 stt.py --train --tb_name libri_shoota_TEDLIUM

Training graphs

The learning rate was initialized at 0.0003. Learning rate

The error rate below is on the test set on which the rnn never train. Error rate on test set

The loss is dropping as expected. Loss