This repository contains the code for the paper "Language Identification Using Deep Convolutional Recurrent Neural Networks", which will be presented at the 24th International Conference on Neural Information Processing (ICONIP 2017).
train.py
and predict.py
.config.yaml
.You can install all python requirements with pip install -r requirements.txt
in the respective folders. You will additionally need to install the following software:
The repository contains a model for 4 languages (English, German, French, Spanish) and a model for 6 languages (English, German, French, Spanish, Chinese, Russian). You can find these models in the folder web-server/model
.
To start a training run, go into the keras
directory, set all the desired properties and hyperparameters in the config.yaml file and train with Keras:
python train.py --config <config.yaml>
To predict a single audio file run:
python predict.py --model <path/to/model> --input <path/to/speech.mp3>
Audio files can be in any format understood by SoX. The pretrained model files need to be caomptible with Keras v1.
To evaluate a trained model you can run:
python evaluate.py --model <path/to/model> --config <config.yaml> --testset True
You can also create a visualisation of the clusters the model is able to produce by using our tsne.py
script:
python tsne.py --model <path/to/model> --config <config.yaml>
In case you are interested in creating a visualization of what kind of patterns excite certain layers the most, you can create such a visualization with the following command:
python visualize_conv.py --model <path/to/model>
0 English,
1 German,
2 French,
3 Spanish,
4 Mandarin Chinese,
5 Russian
GPLv3 see LICENSE
for more information.
If you find this code useful, please cite our paper:
@inproceedings{crnn-lid,
title={Language Identification Using Deep Convolutional Recurrent Neural Networks},
author={Bartz, Christian and Herold, Tom and Yang, Haojin and Meinel, Christoph},
booktitle={International Conference on Neural Information Processing},
pages={880--889},
year={2017},
organization={Springer}
}