Skip to content

Latest commit

 

History

History
 
 

contextnet

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
figs
figs
 
 
README.md
README.md
 
 
config.yml
config.yml
 
 
test.py
test.py
 
 
tflite.py
tflite.py
 
 
train.py
train.py
 
 

README.md

ContextNet: Improving Convolutional Neural Networks for Automatic Speech Recognition with Global Context

Reference: http://arxiv.org/abs/2005.03191

ContextNet Conv Block

ContextNet Se Module

Example Model YAML Config

Go to config.yml

Usage

Training, see python examples/contextnet/train_*.py --help

Testing, see python examples/contextnet/test_*.py --help

TFLite Conversion, see python examples/contextnet/tflite_*.py --help

RNN Transducer Subwords - Results on LibriSpeech

Summary

  • Number of subwords: 1008
  • Maximum length of a subword: 10
  • Subwords corpus: all training sets
  • Number of parameters: 12,075,320
  • Number of epochs: 86
  • Train on: 8 Google Colab TPUs
  • Train hours: 8.375 days uncontinuous (each day I trained 12 epoch because colab only allows 12 hours/day and 1 epoch required 1 hour) => 86 hours continuous (3.58333333 days)

Pretrained and Config, go to drive

Epoch Transducer Loss

subword_contextnet_loss

Epoch Learning Rate

epoch_learning_rate

Error Rates

Test-clean Test batch size Epoch WER (%) CER (%)
Greedy 1 86 10.356436669826508 5.8370333164930344