deepsc-st-pytorch

This repository contains code for JSAC 2021 "Semantic Communication Systems for Speech Transmission". The code is based on pytorch implementation according to the original code and the content of the paper.

Project structureChannel layer

.
├── dataset                 # Dataset
├── channel.py              # Channel layer
├── config.py               # The config of channel
├── custom_dataset.py       
├── log                     # The folder where logs are saved
├── log.py                  # User-defined logger
├── make_dataset_t.py       # Make dataset
├── metrics.py              # Evaluation metrics of the model 
├── model_thesis.py         # model
├── README.md
├── results                 # The folder where model files and evaluation results are saved
│   └── saved_model_thesis_with_noise
│       ├── awgn
│       └── rayleigh
├── speech_processing.py    # Data preprocessing includes normalization, de-normalization, sampling, and de-sampling
├── test_thesis.py          # Model testing
├── train_thesis.py         # Model training
└── utils.py                # Other, include resampler, create folder, ...

Prerequisites

• Python 3.8+ and Conda
• Environment

  conda create -n deepsc_s python=3.8
  conda activate deepsc_s
  python -m pip install -r requirements.txt

Datasets

• The dataset adopted in this project can be found in Edinburgh DataShare, Please download this dataset and put them into ./dataset/raw_data folder. Then do the following:

  mkdir dataset
  python make_dataset_t.py --sr 8000 --num_frame 128 --frame_size 0.016 --stride_size 0.016 --save_path ./dataset/deepsc/

• If you don't want to download the original dataset, you can use the data set we processed directly.

  mkdir dataset
  cd ./dataset
  
  git clone https://github.com/linhl1218/deepsc-dataset.git
  mv deepsc-dataset deepsc
  

Usage

Example of testing model performance change under awgn channel SNR=10dB:

python test_thesis.py --channel_type awgn --snrs [10]

Train the model:

python train_thesis.py --channel_type awgn --num_epochs 400 --batch_size 16 --lr 0.001