Joint End to End Approaches to Improving Far-field Wake-up Keyword Detection

• Python >= 3.7 (Recommend to use Anaconda or Miniconda)
• PyTorch >= 1.3
• NVIDIA GPU + CUDA

1. Install dependent packages

   cd E2E-Keyword-Spotting
   pip install -r requirements.txt

2. Or use conda

   cd E2E-Keyword-Spotting
   conda env create -f environment.yaml

Dataset is from Google Speech Command published in arxiv.

• Data Pre-processing (Has already been done)

1. According to the file, dataset has already been splited into three folders, train, test, and valid.
2. The splited Google Speech Command dataset is saved in Google Drive folder.

• Single GPU Training:

python train.py

• Distributed Training:

CUDA_VISIBLE_DEVICES=0,1 python train.py

python test.py 

1. Multi-head Attention

2. VGG19/VGG16/VGG13/VGG11 with/without batch normalization
3. Deep Residual Neural Network ('resnet18', 'resnet34', 'resnet50')
4. Wide Residual Networks ('wideresnet28_10') imported from the repository
5. Dual Path Networks from arxiv
6. Densely Connected Convolutional Networks from arxiv

├── kws
│   ├── metrics
│   │   ├── fnr_fpr.py
│   │   ├── init.py
│   ├── models
│   │   ├── attention.py
│   │   ├── crnn.py
│   │   ├── densenet.py
│   │   ├── dpn.py
│   │   ├── init.py
│   │   ├── resnet.py
│   │   ├── resnext.py
│   │   ├── treasure_net.py
│   │   ├── vgg.py
│   │   └── wideresnet.py
│   ├── transforms
│   ├── utils.py
├── config.py

• ./kws/metrics : Evaluation matrics, defining the False Rejection Rate (FRR) and False Alarm Rate (FAR) for keyword spotting
• ./kws/models : Diffferent network architecture
• .config.py : Configuration about parameters and hyperparameters